Recent developments in the application of machine-learning towards accelerated predictive multiscale design and additive manufacturing

ABSTRACT The application of three-dimensional (3D) printing/Additive Manufacturing (AM) for developing multi-functional smart/intelligent composite materials is a highly promising area of engineering research. However, there is often no reliable means for predicting and modelling the material performance, and the wide-scale industrial adoption of AM is limited due to factors such as design barriers, limited materials library, processing defects and inconsistency in product quality. A comprehensive framework considering the generalised applicability of ML algorithms at sub-sequent stages of the AM process from the initial design to the post-processing stages in the literature is lacking. In this paper, the integration of various ML applications at various sub-processes is discussed, including pre-processing design stage, parameter optimisation, anomaly detection, in-situ monitoring, and the final post-processing stages. The challenges and potential solutions for standardising these integrated techniques have been identified. The article is promising for professionals and researchers in AM and AI/ML techniques.


Introduction
One of the earliest applications of computers in predicting the behaviour of structural compositions based on fundamental physics (Pople 1999) was by John People-using a programme named Gaussian 70 that could perform ab-initio calculations.His efforts made quantum chemistry accessible to a broader reach of experimental chemists.With the advent of supercomputers and advanced algorithms, more investments and improvements were made in computational investigations in science and engineering.The advancements in computational chemistry and material science involving high-throughput calculations gave researchers the option to have predictive capabilities for developing many compounds as part of a single study.One of the techniques used to calculate structural behaviour is density functional theory (DFT) (Hohenberg and Kohn 1964;Kohn and Sham 1965).This led to the establishment of databases covering known and notional systems, including single crystals, metal alloys (Hachmann et al. 2011;Jain et al. 2013;Calderon et al. 2015), organic and inorganic crystals.
The proliferation of scientific research on advanced smart materials, including advances in computing capabilities, led to the development of several new techniques in the design, manufacturing, testing, and validation fields.Advanced materials can introduce large design spaces owing to their complex topological and compositional structures.Multi-scale modelling is a cornerstone for the relatively new class of hierarchical materials which can perform multifunctional tasks, owing to their electrical, magnetic or thermal properties.The design of smart multifunctional materials necessitates a solution that prevails across structural and functional materials on both the local and the global scale (Lendlein and Trask 2018).It encompasses developing computational design and solution capabilities that enhance the route to novel digital manufacturing paradigms.These advanced numerical tools depend on various experiment-intensive data, giving an understanding of the physics involved across length scales.
Recent advances in Three-Dimensional Printing (3DP) have immensely widened the design space for advanced composites by introducing possible structural hierarchies across scales, novel arrangements of materials, and expanded material selection.However, these may not be suited for conventional design techniques due to current computational limits.Multiscale modelling considers phenomena occurring at various length and time scales and can provide reliable predictive modelling capability in the 'tuning' of the microstructure design process.This can be critical in the distribution of properties across sections of a fabricated part, thereby customising it for specific applications.
3D printing was initially employed to produce prototypes and replicas, including toys, bio-implants, metallic components, test specimens, etc.There are different methods of 3D printing namely material extrusion, material jetting, binder jetting, powder bed fusion (PBF), directed energy deposition (DED), sheet lamination (SL) and vat photopolymerization techniques like stereolithography (SLA).A classification of the various common AM techniques discussed in this study, based on their operating principle is shown in Table 1.The features, such as high customisation and the ability to produce functional 3D structures with control over material properties, have helped 3D Printing of reinforced polymers gain a wide range of interest from various science industries, including automotive, aerospace, biomedical, robotics, and electronics industry to name a few.
With the increasing advances made in one of the lowcost 3DP techniques, known as Fused Deposition modelling (FDM) or Fused Filament Fabrication (FFF) which is a material extrusion technique, 3D printing has now enabled the manufacture of complex multifunctional materials even on the micro/nano-scale.This was not feasible using conventional methods (Frazier 2014;DebRoy et al. 2018;Ngo et al. 2018;Lee, An, and Chua 2017).It could be fully unlocked by multiscale virtual testing tools, which are accurate and efficient.Along with the improvement in printing techniques and materials, there is an emphasis on the creative design of complex systems, as represented in Figure 1, and inspiration can be drawn from several complex biological systems seen in nature that can be adapted into engineering systems.Evolved structures such as bioinspired sensors (Lin et al. 2014), crab claws (Lin et al.  2014), shark skins (Wen, Weaver, and Lauder 2014), etc., provide guidelines for the design and development of complex systems.3DP being a complex manufacturing process, comprising a multitude of influencing parameters and potential fluctuations resulting in a high probability of inconsistencies in properties, defects and imperfections in the fabricated part.This hinders the widescale adoption of AM techniques by various industries and restricts the application of 3D-printed parts.ML can aid in overcoming this limitation by utilising the datasets at various stages of the AM chain to provide optimal solutions, thereby accelerating the transition to a digital manufacturing process.This includes data collected from advanced computational tools such as multi-scale simulations and the application of ML techniques.
Artificial Intelligence (AI) and big data have been popularly referred to as the 'fourth industrial revolution' (Schwab 2015).Its application for virtually all spheres of scientific research has been expanding astoundingly.A subfield of AI gaining prominence and improvement in recent years is Machine Learning (ML).Algorithms, based on statistical techniques, form the base of ML applications.In simple terms, ML can be explained as using computers to learn about a task to completion without requiring explicit programming.ML is already a focus of research in various fields such as mechanical engineering (Geetha and Bridjesh 2020), material science (Juan et al. 2020), construction (Bilal and Oyedele 2020) and medicine (Smiti 2020).ML has the potential to make the field of molecular and material modelling more popular and effortless by lowering the requirements of computer power and prior domain knowledge.There have been various efforts in the past to integrate statistical AI/ML techniques at different stages of part development, from multiscale design to 3D printing process parameter optimisation and in-situ monitoring.Through the various studies discussed in the paper, it can be inferred that integrating ML techniques can improve the reliability of 3D printed parts, and this is vital to prove their utilisation as an endproduct tool.
AM has gained tremendous interest in designing intricate porous polymeric materials due to its unique offerings, such as precise control over spatial porosity during fabrication.This enables the development of highly efficient and customised foam structures, which traditionally required massive investments during development in terms of time and finite element (FE) simulations.Recent research (Roach et al. 2021) has indicated the viability of using artificial neural network (ANN), computer vision algorithm and genetic algorithm to inverse design a foam structure with required mechanical (compressive) characteristics.Furthermore, a cross-sectional image of a direct ink write (DIW) specimen was processed using the framework mentioned above to generate the printing process parameters required to achieve the target response of the specimen for specific applications.This indicates the promising potential for predicting the mechanical response of structures and based on this outcome, an autonomous design of 3D printed specimens using AI/ML techniques.AI/ML techniques, as illustrated in Figure 2, analyse data to perform optimised decision-making and can accelerate the design space exploration by mapping between the part design and performance, thus leading to cost-effective production and large-scale adoption of 3DP.In selecting a product, designers in manufacturing need to consider several indicators like model geometry, size, cost, time, material, etc.The testing of this multi-indicator setting, and the study of their effect can be complex and could be solved efficiently as an optimisation problem by a suitable ML method.3DP can produce highly customisable, high-performance, multifunctional parts, and its full potential could be better realised by applying ML to investigate the process-structure-property-performance (PSPP) relationship.This relation and the effect of ML have not been fully explored yet, especially in 3D metal printing processes such as laser powder bed fusion (L-PBF), a popular metal AM technique owing to its minimal post-processing required and higher resolution capabilities.However, inconsistencies in part quality hinder the wide-scale adoption of L-PBF processes in the industries.In recent studies (S L Sing et al. 2021;Gong et al. 2022), utilising ML to process the datasets at various stages of the L-PBF improved the quality control efficiency, thereby leading to better consistency in quality.The study (S L Sing et al. 2021) emphasised the potential process variability in L-PBF and the requirements introduced by implementing ML algorithms into the workflow-including advanced data tools, data acquisition methods, computer vision and sensor integration for predictive maintenance of equipment.
The ability of some 3DP techniques to print multiple materials has enabled the development of functionally graded structures (Loh et al. 2018).The topological optimisation of lightweight components has also been enabled by the ability of 3DP techniques to manufacture complex geometries.Ensuring the quality and reliability of the products calls for relying on experiments or models entailing the physics of the process.This physics-based computational modelling may involve several limitations, including limited experimental data, physical approximations and model uncertainties.This directs us to the need for an efficient computational effort for combining the experimental and simulation data to produce the desired end-products.
Typical 3DP techniques have many degrees of freedom, as illustrated in Figure 3, of a laser-based AM process and are highly interdisciplinary in nature, with an integrated consideration of materials science, chemistry, physics, electrical engineering, mechanical engineering, and computer science.The application of advanced ML methodology for modelling the mentioned degrees of freedom can heavily reduce the effort in terms of time and cost investment.It can also make the optimisation process more accessible.ML algorithms also have the potential to achieve processcentric objectives by guiding the computational design of 3DP processes.ML can take it beyond the current capabilities, as indicated by the comprehensive report of Integrated Computational Materials Engineering (ICME) (Martukanitz et al. 2014).Both modelling scenarios, either when the model does not exist or when the model is too expensive to simulate by available numerical methods, could have an alternate and more viable solution if ML comes into the picture.For the first case, ML can form a relationship between the output and input from a few measurements (from past simulations/experiments).For the second case, the same previous results (from simulation/experiment) may be drawn to suggest an alternative numerical method.This can help researchers in studying complex relationships involving high-dimensional spaces.
In another interesting study, the efficacy of a datadriven 3DP approach was clearly indicated (Nguyen et al. 2022) for static purpose structures, using a deep learning-based models-Multilayer Perceptron (MLP) viz.an ANN with interconnected neurons and Convolutional Neural Network (CNN)-a MLP with convolutional layers as hidden layers.The ML tool developed was able to optimise the vital printing parameters, such as weight, time duration and length, from a model design irrespective of the shape of the final geometry and external intervention.The study's findings indicated the great potential for improving printing efficiency and reducing the economics involved in the process.However, it requires more study into printing functional parts for fields such as medical and construction.
Current statistical and ML developments do not have the full capability to resolve the challenge of transferring a model across the vast sub-processes and forming a generalised approach for an AM system.There is a need for developing innovative approaches, such as introducing a general framework providing deeper insights into the engineering processes across various sub-processes of the AM system.Advanced developments, including model transfer and active learning techniques, as discussed in the following sections of this article, provide opportunities and potential for a comprehensive modelling scheme connecting the various sub-fields in a distributed 3D printing manufacturing environment.
The paper focuses on evaluating the improvement in process efficiency by applying AI/ML techniques at various stages of the 3DP workflow.As summarised in Figure 4, the main objective is to identify the essential influencing parameters in the process workflow from the design to the post-processing stages, which can further be utilised to build targeted experiments/simulations.It also entails the selection of suitable ML algorithms to extend 3D printing domain expertise and enlarge the training dataset for AI models.In the following section of this review article, several standard ML techniques have been introduced, discussing their applicability at various stages of the 3D printing process.The contemporary ML applications discussed in this paper include multi-scale modelling and the challenges in resolving the molecular-and micro-level mechanisms guiding the macroscopic material behaviour, model design for additive manufacturing (DfAM), parameter optimisation, in-situ monitoring in the area of interest during the process, part property prediction, quality assessment and control.Finally, the challenges, research gaps and future research directions on ML applications in 3DP have been highlighted.

ML-based surrogate models for 3D printing
The integration of 3D printing with neural networks can enable the autonomous production of test specimens and data feedback for model improvement, thereby enhancing predictive capabilities.This may involve advanced robotics for manufacturing (Nikolaev et al. 2016), testing and material characterisation.The ML algorithms applicable for various 3DP can be generally categorised into mesoscale (e.g.melt pool geometries) and macro-scale level (e.g.mechanical properties).Each scale mentioned above may affect the quality of the final fabricated part, indicating ML to be more beneficial.Topology optimisation with given constraints using ML capabilities can also provide new opportunities for higher efficiency, optimisation, and savings (Mozumder, Mairpady, and Mourad 2019;Idrisi and Mourad 2019;Yousef, Mourad, and Hilal-Alnaqbi 2011;Thekkuden and Mourad 2019) and better performance even at the nano (Chen, Chrzan, and Gu 2020) and continuum (Rozvany 2001;Brackett, Ashcroft, and Hague 2011) scales.

Types of ML algorithms and approaches
A computational model based on ML, popularly known as a surrogate model, can have different algorithms based on algorithm architecture, available size of the dataset and computational power required.It comes down to the end-functionality of the model and associated proven characteristics.Depending on the mode of operation, ML algorithms are classified as supervised and unsupervised learning.A classification scheme is depicted in Figure 5, indicating various types and examples of popular ML algorithms.A generalised application for these different categories of algorithms classified based on their operating principle is also depicted in Figure 6 for various AM-process sub-fields.The supervised learning approach adopts a function that maps labelled training datasets as input and later can be used to predict the label of unlabelled cases.The ML algorithm is used to infer a function based on the training data and map it onto new output data.Supervised learning problems are categorised into regression and classification.Regression deals with quantitative labels, such as the estimation of instances of an object in a given image (Friedman, Hastie, and Tibshirani 2001).Classification deals with qualitative data, such as to indicate if a given image is of a particular kind or not.The most popular learning algorithms include Linear regression, Logistic Regression, Neural networks, Multilayer perceptron, Support vector machines, Naïve Bayes, Decision trees and the Knearest neighbour algorithm.
One of the most popular tools mentioned above is Artificial Neural network (ANN) mainly because of its compatibility with handling large datasets and efficient inference of non-linear decision boundaries.However, as illustrated in Figure 7, the mapping between input and output spaces may be of different dimensions and consist of a series of processing units known as hidden layers.Also, specialised CNNs are developed for specific applications such as computer vision.CNNs explore the input space to identify similar features, such as vertical lines in computer vision tasks.Also, neural networks with more than two hidden layers are classified as deep learning (DNN).Hidden layers multiply the weight factors to the input layer and add some bias related to the output layer.
Another powerful and flexible supervised ML algorithm is the support vector machine (SVM) which can be used for regression and classification.However, it is popularly used for cases involving classification.SVMs find the maximum marginal hyperplane (MMP) by splitting the datasets into classes (Friedman, Hastie, and Tibshirani 2001).Hyperplanes are iteratively generated, and the one that divides the classes accurately is chosen.A kernel function may be used to implement the SVM algorithm, effectively transforming the input data space into a higher dimensional or required form.
Similarly, data is best fitted for the regression tasks using the hyperplane.Although it is proven that SVMs are suitable for high-dimensional space, they need higher training time for large datasets.Furthermore, the chances of overfitting are higher in features than in training examples.
In unsupervised learning, patterns are discovered from unlabelled data.No manual intervention is required to supervise the model, although this method can be unpredictable comparatively.There are various types of unsupervised learning-based algorithms, the prominent ones being association rules and clustering.Association rules aid in identifying relationships between different variables in large datasets (Friedman, Hastie, and Tibshirani 2001), such as those associated with image recognition, etc. Clustering/data segmentation identifies a pattern in the uncategorised data and can group them into different clusters (Friedman, Hastie, and Tibshirani 2001).
Reinforcement learning is another cutting-edge technology developed having massive potential in 3DP.In this goal-oriented approach, the ML model is trained for performing sequential decision-making through the optimisation of a reward signal (Sutton and Barto 2018).The computational power of computers to perform trial-and-errors is leveraged to maximise the reward signal (Sutton and Barto 2018).Reinforcement learning differs from supervised approaches as there are no training scenarios for securing the reward.Also, it is different from the unsupervised learning approach, as there is no attempt to identify patterns within the datasets.
It is possible to develop surrogate models based on any of the commonly used ML techniques mentioned above.However, attention is required in choosing the candidate based on factors such as interpretability suited for the purpose.The user must be aware of potential pitfalls in the form of recording errors and biases in data and learning.Factors such as feature selection, feature extraction, feature engineering and dimension reduction affect the interpretability of ML approaches.Inconsistencies and biases are often inherited and magnified by surrogate models.The case studies at each subfield of AM process workflow, described in the upcoming respective sections, illustrate the operating principles and structures of the various specialised ML algorithms and their suitability, indicated by resultant output and comparative analysis.

Algorithm-driven 3DP
The ability of ML-based algorithms to learn principles and detect patterns behind datasets attracts interest and makes them a good candidate for application in the 3DP systems.The 3DP workflow sequence, as indicated in Figure 8, involves several process parameters, and their combination affects the quality of the final product.This can often induce errors in the process and may depend on the expertise of the operator of the 3DP equipment.The detection of these errors and subsequent correction can affect the reliability and efficiency of the process.Several studies (Jin, Zhang, and Gu 2019;Gardner et al. 2019) have been made leveraging the advantages of ML to study the intrinsic relationship between product quality and process parameters.

Building material datasets
The basic layout of a ML process, as illustrated in Figure 9, includes preparation (also termed as cleaning) of the input data from multiple sources, selection of suitable descriptors for appropriate representation, selection of algorithm, and finally, using the developed model for prediction and further applications (Liu et al. 2020).Cleaning data is essential to eliminate incorrect information, and subsequent steps include normalisation, standardisation and stratified sampling, which handles imbalance.Any data changes also necessitate optimisation of related hyper-parameters.Unseen data should be used to test and cross-validate models, compare results with experimental results and draw inferences based on domain knowledge.Data fed as input to the ML model of choice can be sourced from existing databases, literature or high throughput simulations and experiments.The data creates the relationship between the applied ML model to the mechanics of the problem in the investigation.The main challenges in this aspect are ensuring that the right amount of data is obtained and that its quality is such that the efficiency of the ML model is maximised.
The recent interest in computational material design has led to the creation of many material databases like MatMatch ('MatMatch'), Materials Project (MP) (Jain et al. 2013), MatWeb ('MatWeb') and MatNavi ('MatNavi'), to name a few.Material properties, including elastic constants, tensile, shear, flexural, fatigue strengths, hardness, etc., may be derived from theoretical calculations and experimental measurements.Researchers have the freedom to balance the efficiency of the ML model and the cost of the required data generation.While performing simulations or experiments, there is more space to control the size and distribution of expected datasets.
Adequate domain knowledge of researchers must be exploited to perform data pre-processing and obtain representative data.There are several cases where the ML model would fail or seem unnecessary, including the availability of alternative methods to explore the design space at a relatively acceptable cost.Also, the non-availability of datasets covering the whole design space and the challenges faced in the interpretability of datasets collected may cause difficulties in framing a ML model for a specific application.

Data representation
The initial raw data, which may or may not be numerical, needs to be converted to a convenient form for an algorithm to process through a method known as featurization/feature engineering.The ideal representation of input data can affect the algorithm's effectiveness in mapping it to output data, mainly supervised learning algorithms.The choice of representation can give an idea of the underlying algorithm and the scientific problem in the study.It is still an active field of research (Faber et al. 2017) to obtain the right choice of representation such that overall performance is upgraded.
The numerical representation of input cases or materials can be termed fingerprint, feature vector, or descriptor.Depending on the accuracy and level of output property under study, the fingerprint can be defined (Ramprasad et al. 2017) at various scales of granularity, right from a higher-scale level (grain size) to sub-Angstrom resolution cases (Mueller, Kusne, and Ramprasad 2016;Friedman, Hastie, and Tibshirani 2001).In addition, techniques such as principal component analysis (PCA) (Rajan 2005) effectively search for descriptors in multivariate data, reducing the information's dimensionality.It is beneficial for cases of large databases, simulations or experiments, where phenomenological relationships exist and may not be possible to be explained apriori.
Scientific knowledge can be encoded at the preprocessing data stage and in the architecture of the ML algorithm.One such example, a physics-informed neural network (PINN) (Raissi, Perdikaris, and Karniadakis 2019), has huge potential to reduce the computational cost of material behaviour analysis.

Applications of ML at various stages of additive manufacturing
Scientists and researchers should ensure that they comprehend the framework and methodology of ML, which operates on mathematical operations.These ML tools, similar to DFT or FEM simulations, have their capabilities as well as limitations.For example, data lacking statistical correlation is not suitable for the applicability of ML algorithms.Therefore, the foremost concern of the scientists and researchers is evaluating the data to decide if the ML technique is appropriate enough and meaningful if applied.One such case was pointed out (Johnson et al. 2020), where the location of maximum pore size is to be predicted in a L-PBF build-up process.The fact that ML is not suitable for this can be concluded from observing that the maximum pore size (even if it is repeatable, i.e. occurs at the exact location for multiple build-ups), being a statistical outlier, would be ignored by all ML algorithms.Therefore, this understanding is essential before working with predictive ML models and their application.Although this study is focused on the design and engineering aspects of 3DP techniques, the science and technology aspects are also now being widely researched (Wang et al. 2020a).It includes applications such as new alloy design, topology optimisation, microstructure characterisation and property prediction.
An overview of the applicability of ML at each stage of the process sequence of 3DP, as illustrated in Figure 10, will be discussed going forward.Several important case studies regarding ML applications in various sub-fields of the 3DP techniques discussed in the paper is also compiled in Table 2.

Multi-scale modelling approaches
Developing an accurate and reliable multi-scale modelling and simulation framework is vital to enable effective design predictions and process optimisation.The wide-scale adoption of 3D printing by various industries like automotive and aerospace envisages an integrated material and structural tool, leading to innovative designs with considerable weight and cost savings.However, experimental feedback will not be sufficient to maintain a robust production workflow.
The new 'materials by design' approach has multiscale modelling and simulation as a core component.Multi-scale modelling enables a holistic integration of engineering and scientific methodologies and knowledge, including the unification of material information to performance analysis and process simulation of product manufacturing.A multiscale model is one in which information across various scales, as shown in Figure 11, is considered to comprehend the system better, thereby reducing the overall computational cost.A successful multiscale model must overcome several challenges (LLorca et al. 2013), primarily the balance between accuracy and computational cost.Multiscale models are the most vital component of the socalled Integrated Computational Materials Engineering (ICME) paradigm.It is based on the Process-Structure-Properties Performance principle and encompasses a study of several aspects of a material design cycle, including processing, arrangement of internal structure, properties and performance.Computational techniques such as molecular modelling and ML are creating revolutions in the material design paradigm.It is now possible to obtain a microstructure according to the user's desired properties.
There is no single multiscale theory/experiment spanning time and length scales in assessing the structureproperty relationship, which can often be non-linear.However, approaches such as 'material informatics' (Rajan 2005) have been developed, aiding in material discovery and design.The approach provides insights by analysing material data at scale.A bottom-up approach in designing composites right from the nanoscale to higher has been made by several researchers, as shown by the basic illustration in Figure 12, where molecular dynamics (MD) is used to simulate the characteristics of the matrix and fibre-matrix interface and upscaled.Such multiscale simulations can be evaluated at multiple scales using stress transfer behaviour, and this can aid the designer in fine-tuning the desired material  behaviour.A ML-based interatomic potential was also shown (Mortazavi et al. 2020) to perform first-principlesbased multi-scale modelling of thermal conductivity for borophene/graphene lattice heterostructures.
ML and numerical methods both involve models in their methodology.However, in ML, the inputs are known prior to training, and the model is unknown, whereas, in numerical simulation, the inputs can be variable and unknown even though the model is known.A review can be made (Nasiri and Khosravani 2021) where the above features can be utilised to provide applications for their area of intersection.This integration can lead to faster cycles of optimisation processes.
. Integrating ML into simulation: The application of ML in different components of simulations can, in effect, reduce the order of the model.The developed surrogate model may not offer exact solutions, although solutions are generally more straightforward.Parametric study of simulation results is one area explored using an integration of ML into the simulation process.The impact on the overall product and engineering process as a function of combining ML and finite element data (FEM) was analysed (Bohn et al. 2013), and a significant reduction in data analysis efforts was recorded.It was demonstrated how ML could impact the design process of new car models by reducing parameters to be investigated and identifying non-influential variables.Methods such as dimensionality reduction and clustering facilitate the division of simulation into groups based on the desired output behaviour.Subsequent processing by suitable reconstruction algorithms can eliminate the need for full numerical simulations under the required parameter configuration.Such data analysis division in the workflow and handling bundles of simulation runs can drastically reduce the run-time invested by the designers and engineers on the analysis and parametric study towards an objective function (reducing cost, weight, desired properties, etc.) in the process of product development.The impact is also evident in fields like numerical weather predictions (Bauer, Thorpe, and Brunet 2015), comparable to any computationally intensive work.The developed model utilised modern hardware and deep learning models and could produce accurate predictions in 4-7 units of time, corresponding to almost 30 years of progress.Several recent studies have reported the viability of integrating ML techniques with FE simulations.A highthroughput multiscale framework can provide consistent, balanced, complete and representative data, which are prerequisites for an accurate ML model.In one exciting work (Lyngdoh and Das 2021), the strain sensing capacity and the electromechanical response of a smart cementitious composite were determined.As outlined in Figure 13, a Shapley Additive Explanations (SHAP) algorithm was used to make a comparative study in estimating the most vital design parameters like coating thickness, etc., affecting the sensing ability of a smart cementitious composite.It may provide reliable standards in the design of composites for structural health monitoring applications.Breuer and Stommel (2021) utilised 3D representative volume elements (RVEs) for finite element simulations of a short fibre reinforced polymer and used them as a training database for an artificial neural network (ANN) to predict the elastic properties.The accuracy of the model predictions was verified for various values of the design parameters of the composite.The results indicated that costly computational effort could be saved, especially for the different combinations of representative volume elements (RVEs).Also, it indicated that the passage of information across scales, e.g. from micro-to macro-scale, could be performed with comparatively lesser computational time investment at the macro-scale level without the loss of predictive capability.
Yang et al. ( 2019) have shown how robust coupling ML methods and high-throughput molecular dynamics (MD) can help predict silicate glass stiffness values.A comparative study was also done among selected ML algorithms such as Artificial Neural Network (ANN), random forest (RF) and polynomial regression (PR).It was found that ANN offered the most accuracy, although the level of interpretability and simplicity varies for each.
In one study (Revi et al. 2021), a binary alloy's elastic properties (Young's modulus, bulk modulus, shear modulus and Poisson's ratio) were derived using a database of DFT-computed elastic constants and compositionally-averaged atomic or bulk elemental properties, to develop the ML model.The accuracy of the developed model was validated by using experimental values of elastic constants as test data, and its capacity for predicting properties of five-component alloy composition was also explored.Another similar research (Kim et al. 2018) illustrated how a ML model could be used to predict the properties of polymer material from past data of polymers belonging to the similar chemical class and properties like refractive index, density, bandgap, dielectric constant, solubility and glass transition temperature, which are dependent on length-scale features.An optimised set of features was used for each property, as different properties are a function of the specific length-scale feature.In addition, a 'fingerprinting' scheme was used to numerically represent the features of the polymer right from the atomistic to morphological structural level.All the studies mentioned in the section highlights the progress achieved in the predictive capability of ML models for various case studies.This strengthens the view to invest into more research in this field that would lead to further development.
A generalised multiscale finite element model based on a method named FE 2 was developed (Feyel 2003), which has the advantage of not requiring macroscale level constitutive equations and having all non-linearities come from the microscale level.The assumption of scale separation was discarded by the FE 2 method when it was used to couple the generalised continuum at the macroscale and classical continuum at the microscale level.The behaviour of a highly heterogenous structure was modelled using classical description at the microscale level and a generalised description at the macroscale level.In one stochastic-based application (Lu et al. 2021), the FE 2 method was used to substitute the non-linear multiscale calculations with the hybrid neural network-interpolation (NN-I) scheme in order to solve uncertainties at both the micro-and macro-scale level.A significant reduction in computational time was recorded, which facilitates Monte Carlo simulations of nonlinear heterogeneous structures, and a study on the electric conduction of graphene-polymer composites was performed.At the micro-level, the data obtained from the finite element simulation on RVEs were used to train the neural network model.This method can take into account the non-linearities at micro-level, which are often ignored otherwise, thereby improving the accuracy and efficiency of the modelling process.
The inherent multi-scale nature of advanced hierarchical structures calls for a multi-scale analysis to study the various underlying mechanisms.Furthermore, advanced manufacturing tools like 3D printing need to be explored in support of multiscale computational modelling for improving the development of novel advanced materials.There are also promising opportunities for inducting AI/ML into the various stages of design and manufacturing advanced smart materials, including advanced material modelling as discussed above and modern techniques of 3D printing.A robust multiscale framework can provide complete, balanced, consistent, and representative data, which can serve as a base for an accurate ML model.

Challenges in multi-scale model functionality
The modelling of the various simultaneous sub-processes as part of the 3D Printing process involves multi-scale problem evaluation.There has to be consistent control over the design of a nanocomposite structure so that properties of the nanofillers like graphene are translated into the nano-, micro-and macro-scales (Ponnamma et al. 2020).There is a need for further investigation into factors affecting the process, like material parameters and manufacturing processes, to obtain a sustainable design and manufacturing cycle.It has been observed that there are mismatches between the structural scale and microscale properties.These are reflected in the failure strengths, stiffness values (Sochi 2012), and nano-composite characteristics that do not scale up linearly.
Defining process-structure-property-performance relationships of materials based on physics-based and date-driven mathematical relationships is vital in assessing the reliability aspect of material design.The current modelling paradigm is based on a 'bottom-up' approach, which utilises single-scale methods, and passes these parameters across scales.Traditional macroscopic continuum methods such as finite element, finite difference, and computational fluid dynamics do not offer spatially resolved information and are based on mean-field methods or equilibrium conditions.Current commercial solutions use constitutive model formulation to represent material at an integration point within the FE method.Enhanced capabilities such as the UMAT routine (ABAQUS) enable user-definable constitutive and failure models within commercial codes.Olsen devised a design concept, as shown in Figure 14, connecting the process-structure-property-performance relationships and a route map to produce highly customised and advanced materials with intended properties.
A first step in understanding smart nanocomposite behaviour at higher scales necessitates a methodical comprehension of the load transfer mechanics between the matrix, nanofiller and microfiber.The design of interfaces can be utilised as the foundation for the creation of novel hierarchical composites with unprecedented physical and mechanical behaviour.The micromechanical interface interaction between the nanomaterial and polymers, especially at length scales of nanometre-to-micrometre, is an active area of research and is vital for in-situ measurements (Buehler and Misra 2019), which are essential in the self-sensing capabilities of high-performance smart composites.Novel innovations such as in-situ curing and local microwave-based welding of 3D printing carbon-filled polymer apply to nano-composites based on graphene (Sweeney et al. 2017).Moreover, field-assisted techniques such as electrical and magnetic are breakthroughs for controlling the interface property of each layer of a printed structure (Yang et al. 2018).
Recent studies involving CNT nano-composites reveal that a bifurcation exists at higher scales for predicted performance due to the unique stress state at the nano-scale (Subramanian, Rai, and Chattopadhyay 2015).Furthermore, from atomistic studies, it was observed that bond breakage and formation might affect the macro-scale elastic properties and damage modes (Rai and Chattopadhyay 2018), which may not be accounted in conventional continuum techniques.Extensive experimental characterisation (Pal and Kumar 2016;Alam et al. 2019) is one approach, although the number of complex interactions and physical variables can make it cumbersome.Another approach is the development of advanced computational tools, integrating nanoscale mechanics to resolve atomistic simulation to find perception into the stress transfer and fracture mechanics at the structural level.
Hybrid models are to be developed based on multiscale modelling approaches, in which several categories of atomistic methods (such as molecular dynamic [MD] or molecular static [MS]) are coupled with discretized continuum methods (FEM).A critical area would be how communication between these atomic and continuum regions is handled.Complications observed include impedance mismatch between finer and coarse regions (leading to spurious wave reflection), ghost forces from the coupling of systems and entropy-like contributions, which can affect the dynamics of coarse-scale models (Liu et al. 2018) caused due to the thermal motion of atoms.
Previous studies have indicated that multiscale modelling and ML techniques can complement each other to face complex problems, manage noisy data and expand 'massive design spaces' (Alber et al. 2019) by enhanced physics-integrated predictive models.The fact that ML models can learn from multiple data sources, like experiments and multiscale models, the idea that ML models may be superior to data sources, is widely posed.However, multiscale modelling alone can often not effectively combine large datasets originating from various sources and varying levels of resolution.It has also been suggested that ML approaches can be beneficial for linking models across scales in cases where connecting through other methods is challenging.However, such hybrid approaches need to be thoroughly examined to eliminate the chances of flaws or biases which may be inherently related to the physical data or ML technique.

Material and geometric design
Various design variables have been fed as input parameters/features to ML models, and they have been utilised to predict various chemical and mechanical properties such as stability, toxicity, deformation, stiffness and strength (Wanigasekara et al. 2019;Wagner et al. 2019;Yang et al. 2018;Sherman, Simmons, and Przybyla 2019).One such study (Hashemi, Safdari, and Sheidaei 2021) was made using a supervised ML approach based on the structure-property relationship for predicting the thermal conductivity of particulate multifunctional composite materials.It was demonstrated that the design process of composites with targeted properties could be accelerated more effectively than conventional optimisation techniques.The role of choice of physical descriptors forms the base of how complex the modelling and establishment of linkages are going to be.The surrogate model was then trained and used in the inverse design of a new category of composites termed as liquid metal (LM) elastomers.
A design for additive manufacturing (DfAM) framework integrated with ML was proposed (Jiang, Xiong, et al. 2020), wherein process-structure-property relationship analysis could be carried out in both directions.An example was highlighted in designing and fabricating a tunable stiffness ankle brace using material jetting 3DP.A deep neural network (DNN) took the structure-property input-output data to compute structural design variables without the need to invert the relationship.This was unlike Gaussian process regression (GPR) and other conventional surrogate models, which need an optimisation method for inverting to estimate design variables.The DNN model was more accurate and computationally efficient than the GPR model.
The choice of an environmentally sustainable option for a designer is a topic of several studies.(Yang, He, and Li 2020).Even though earlier studies report the environmental impact of 3DP in the production stage by correlating the process parameters and energy consumption, this work was one of the first to evaluate the effect of the geometry of each printed layer, i.e. shape and surface area, on energy consumed.Power consumption is a significant factor in a 3DP process like mask image projection SLA (MIP-SLA), which is classified as a vat photopolymerization method according to the ASTM standards.Three ML models, namely NN, stacked autoencoders (SAE) and linear regression (LR), were used for a comparative study and SAE was found to have the best testing performance with 0.85% root-mean-square error (RMSE).The results provide 3DP designers with layer-wise information well before fabrication, thus enabling optimisation for efficient and cleaner production.A hybrid ML approach integrating DBSCAN, which clusters unstructured data and eXtreme gradient boosting (XGB), was similarly adopted (Li et al. 2021) to predict energy consumption in a selective laser sintering (SLS) process, which is another form of the ASTM standard L-PBF, having several subsystems and influential factors affecting energy consumption.
It is critical that the given design meets the requirements of performance and other optimum standards.The ability to produce complex structural configurations with customised material distribution is one major advantage of the 3DP process.Research is still underway to make this process even smoother by lessening the dependency of the 3DP on the overhang and additional support structures, which are usually required to conduct the process reliably.Topology optimisation (TO), referred to as the efficient distribution of material within given constraints has attracted interest since the late twentieth century.
One of the significant concerns to be investigated within the context of 3DP is the overhangs and the related support structures, which can make the TO process challenging.One interesting feature of fabrication based on TO in 3DP is that both bulk and intricate lattice structures have the same time efficiency during operation.During fabrication, any non-existing regions can be customised for any optimised design, as required.This approach was used in the structural optimisation of a cooling channel (Cheng et al. 2018), producing graded lattice structures.This approach was later leveraged for multi-scale problems (Gao, Luo, Xia, et al. 2019;Sivapuram, Dunning, and Kim 2016;Gao, Luo, Li, et al. 2019), wherein microstructures and global geometry are optimised jointly.However, in cases where multimaterial or multi-scale designs exist, there are several challenges to be overcome, especially at the phase transition boundaries.Experimental constraints observed (Holland 1992;Jin, He, and Du 2017;Fera et al. 2018;Zhang et al. 2017) at such interfaces include heterogeneities, lack of fusion, gradient properties and variation in strength characteristics.These may be specific to each 3DP process, the material used and process parameters and this need to be incorporated into the design phase to obtain the desired properties of the final part.In one promising study (Abueidda et al. 2021), a temporal convolutional network (TCN) and gated recurrent unit (GRU) were shown to accurately predict the history-dependent responses, which are dependent on phase transformation, time, temperature and loading path for a thermo-viscoplastic solidification model.The scope of ML in the management of tolerancing issues, including geometrical shape deviations, was indicated (Zhu et al. 2018) using a Bayesian inference-based predictive model.This reinforces the reliability of results obtained from predictive models.
The design freedom offered by AM techniques can be utilised in the design phase of the process chain.For example, the design for additive manufacturing (DfAM) concept involves matching several considerations such as microstructure, material composition, part shape and size to the manufacturing capabilities of the AM process.ML can play a vital role in achieving these objectives, including optimisation efforts such as weight savings.ML models also have the capability to consider the design and process factors and determine the manufacturability of a part.

Optimisation of process parameters
Traditional approaches towards predicting product quality, including trial and error experimental techniques, may prove to be expensive and time-consuming, depending on the scale.In addition, the operator may also need to validate the range of process parameters before the final set-up is confirmed.In this context, data-driven ML approaches have the potential to enhance quality, reliability, cost and time savings.
Depending on the type of the 3DP process, various factors govern the quality of the final product.ML models can be trained to study this relationship more in-depth, derive predictions, and accelerate the optimisation process.For example, laser-assisted 3DP processes such as L-PBF have critical parameters such as laser speed and power.In the case of polymers, where FFF is mainly employed, the primary operating parameters include printing speed, individual layer height and material flow rate.There is also a need to understand better the underlying physics involved in the printing process, which can clarify the interdependencies between parameters, and ML can provide insights into unravelling them.
A Gaussian process regression (GPR) was also utilised (Gong et al. 2020) to draw process-microstructure-properties correlations in the Laser Engineered Net Shaping (LENS TM ), which is a proprietary name for the ASTM standard term-DED additive manufacturing technique (Ang, Sing, and Lim 2022), of compositionally graded Ti-Mn alloys.The DED technique melts metal powders onto a substrate using a high-energy laser beam.An in-depth analysis of data from the above study indicated several reliable correlations, including ageing characteristics to varying Mn content and microstructure parameters.
There may be different applications of ML techniques depending on the desired output.For example, in one study (Gardner et al. 2019), a CNN network was employed to detect the presence of flaws in the final print product of a FDM process for different combinations of process parameters.Optimisation may be done by deriving the ideal parameter combination, which produces zero flaws.Another work (Chowdhury and Anand 2016) involved applying a feedforward neural network (NN) to dampen dimensional inaccuracies caused due to residual stresses, which limits the full potential of 3DP.As illustrated in Figure 15, the NN is initially trained by feeding the simulations of postdeformed geometries.The next step involved providing as input a part geometry into the NN and a compensated geometry is obtained, which is post-processed to derive the final printing geometry.
ML-integrated printing path planning strategies were discussed (Jiang and Ma 2020) based on the principles of a study (Liu et al. 2019) on the deployment of ML in driving cars for selecting the shortest path.With the start and end points being well defined, ML can efficiently produce the ideal print sequence and path by dividing the path into distinct points, and the order of printing is decided through ML.Such path planning strategies provide huge potential in improving quality, accuracy, time, and achieving customised properties of the part and less wastage of feed material.
The effect of varying combinations of process parameters like extrusion speed, print speed, layer height, path distance, etc., in a FDM technique, can be monitored by analysing the nature of the resultant connection between paths in a layer.A detailed study (Jiang, Yu, et al. 2020) was conducted in this field, where almost 400 experimental data was used to train a DNN model for predicting the connection between paths in different scenarios of process parameter combinations.Cases in which overflow across layers occur cause weakness and loss of structural integrity of the final product.With a prediction accuracy of above 80% for connection status, the proposed model could well be developed to achieve beyond dimensional accuracy of the part (connection of path here) to the investigation of material properties, thereby guiding the selection of the ideal process parameters to achieve specific objectives.
DED technologies along with PBF form almost 90% (Gong et al. 2022) of the production-grade metal 3DP systems in operation worldwide.Although the operation of both the techniques is similar in several aspects like layer-by-layer and line-by-line manufacturing, DED offers an upper hand owing to features like higher energy laser input facilitating homogeneous reinforcement distribution, better flow and larger size of the laser molten pool (Chen et al. 2022).A combination of recurrent neural network (RNN)-Deep Neural Network (DNN) was utilised (Ren et al. 2020), to predict, with more than 95% accuracy, the thermal field in a DED process for different geometry and laser scanning strategy in a single-layer build.The above study formed a data structure out of the deposition status.The temperature field was initially built from a FE model that monitored the thermal field evolution during the DED process.This relation between the scanning pattern and temperature field was used to train the RNN-DNN model.Thermal field analysis is vital in the study, as well as the optimisation of the residual stress and distortion commonly encountered in the DED process.
A probabilistic approach-based ML technique was developed (Pandita et al. 2022) for experimental design and uncertainty quantification (UQ) in a meltpool model of a L-PBF process.The initial probabilistic method employed for the study was a multi-fidelity probabilistic modelling-based Bayesian technique (Kennedy and O'Hagan 2000), which combines experimental data with simulation data.
The uncertainty information obtained from the model can reduce the data points of experiments and simulations and improve the model quality.This uncertainty quantification approach has been shown to have advantages, especially in cases where data has been sparse, which may occur due to several reasons.First, a predictive model of the experimental parameter values of melt-pool depth and width was framed from an initial experimental and simulation data set.Then, the meltpool model inconsistencies with the variation of the L-PBF parameters can be analysed.The following method involved applying deep neural networks for transfer learning (Goswami et al. 2020;De et al. 2020;Kaya and Hajimirza 2019).It involved capturing information in the form of physics from one point to another in the process flowchart.The data from the multi-fidelity model after experimental validation and, if above a given threshold of accuracy, is fed to train the parameters of the deep network, termed a probabilistic deep neural network (PDNN).A couple of case studies were also described indicating how this probabilistic deep-based framework can improve highfidelity predictions of melt-pool and lower the number of model simulations and experiments necessary.A similar study in L-PBF (Wang 2020) outlined the Uncertainty Quantification (UQ) and its mitigation using a combination of standard experimental data and high throughput model simulations powering a surrogate model.A Bayesian calibration scheme was selected to process model corrections and calibration of experimental parameters.The impact of this Uncertainty quantification (UQ) framework and its potential role in quality control was demonstrated by predicting lack-of-fusion porosity using a probabilistic porosity model (Tang, Pistorius, and Beuth 2017).The two factors influencing the predictive ability of the surrogate model are the correction required for the model bias and the calibration of sources of uncertainty parameters.
An investigation was made.(Huang and Li 2021) to determine the effect of downstream production parameters such as post-chamber pressure drop and location of the part on the build plate in a L-PBF process on the static mechanical properties of the part.The work aimed to ensure quality repeatability and quality control using an ML-guided framework to identify influential production parameters.This study is distinct from previous ones in that the laser-related process parameters were treated as fixed here.In contrast, others focused on the effects of process parameters on part properties.Several classifications of ML models were deployed for this case to make a comparative analysis, and tree models like RF and Decision Tree (DT) were found to perform best.The results obtained from the parametric study were tested in a follow-up experiment, and the properties were positively enhanced, indicating the potential for quality control and repeatability.

Optimisation of process parameters in the FDM processa case study of process-structureproperty relationships
The recent advancements made in 3DP show immense potential for producing advanced functional, customisable materials with desired performance, cost, and lead time savings.Advanced functional materials based on reinforced polymer-based systems show good manufacturability using the popular 3DP technique, FDM.Such systems have distinct characteristics from the interaction of phenomena at different time and length scales.Processes like FDM have various influential process parameters, as illustrated in Figure 16, which results in highly interdependent sets of final parameters that will dictate the quality of the printed parts and their effective properties.The properties of FDM products are governed by the combined effect/control of various process parameters, structure and material composition.The key to determining the material's functionality is its composition (Reddy et al. 2018;Yu et al. 2017), which governs the heat resistance, thermal effects, melt flow property, etc.Further development and manufacturing of such reinforced-polymer materials using FDM or any other 3D Printing technique, in that matter, necessitates a process control to meet the structural and dimensional requirements for the developed part design.
The application of advanced AI/ML capabilities to digitalised processes such as FDM and other 3DP techniques can improve the product quality and productivity of the process immensely, thereby promoting it to be commercially more viable and for wide-scale adoption.However, the integration of AI/ML across the manufacturing process domain will need a close-up study of the various parameters affecting the finished part microstructure and its performance.In the subsequent paragraphs, several case studies on how these factors affect the performance and multi-functionality of the printed parts will be evaluated, thereby guiding in the establishment of a resilient production workflow.
FDM generally involves material mechanics, fluid mechanics, rheology, heat transportation, phase transition, etc., and each can govern the printability, properties and application of the finished FDM products.Several guidelines have been introduced, such as dealing with stress concentrations at corners (Ismail Mourad et al. 2019;Al Jassmi, Al Najjar, and Mourad 2018) and smaller bead width to ensure better surface quality.In addition, the influence of build orientation on the part accuracy and building parts ensures tensile loads are carried axially along the direction of printing (Parandoush and Lin 2017).Figure 17 gives a schematic of the controlled extrusion process involved in FDM, where a layer-by-layer build-up approach is employed.
Rapid prototyping and manufacturing (RP&M) represent a particular class of fabrication techniques based on a layer-by-layer and point-by-point build-up technique.Earlier works concentrated on 3D printed products with high structural integrity (Zhang 1997;Manthiram, Bourell, and Marcus 1993;Renault 1994).In one of the earliest works (Krope et al. 2017), the feasibility of using FDM to produce short glass fibre-enforced composite with Acrylonitrile-butadiene-styrene (ABS) was demonstrated.The raw materials were combined in a twin-screw extruder to form pellets, which were then drawn into a filament.Improvements in tensile and adhesive strength with increasing Glass Fibre (GF) content (up to 30%) were recorded at the expense of reduced flexibility.
It is vital to enhance the properties of various thermoplastics used in 3D printing by applying nanotechnology.The synergy of 3D printing and nanotechnology can open up opportunities to produce engineered materials with optimised properties, including multi-functionality (mechanical, thermal, electrical, etc.).With the current focus on nanotechnology, it has now become feasible to produce hierarchal-structured polymer composites, requiring the need to process a variety of fillers-including Carbon Nanotubes (CNTs), Carbon Microfibres (CMFs), Carbon Nanofibres (CNFs), graphene-based fillers, etc. to improve specific material performance.Figure 18 shows the microstructure of FDM-produced polyurethane (PU) embedded with nano-additives.
Reinforcing fillers to polymers in 3DP can not only be used to improve mechanical properties but also control the structural transformation in smart composites.Recent advances in 3DP have made precise material placement even at micro-scale possible, opening a field popularly known as 4D printing where the stimuliresponsive printing material can transform shape under changes in temperature, light, or exposure to liquid.These programmable structures can transform from one-or two-dimensional structures to 3D shapes.A comprehensive review of thermally actuated FDM printed PLA, HIPS, and ABS parts was published (Rajkumar and Shanmugam 2018), showing how shape transformation can be controlled for morphing and curved structure applications.A thermo-viscoelastic-viscoplastic model was used to analyze the mechanics involved in the shape transformation mechanics, viz., strain release and storage mechanisms.
Several groups have been actively studying the sensing capabilities of nano-enabled structural materials (De Villoria et al. 2011), where advanced composites containing electrically conductive aligned carbon nanotubes (CNTs) are employed, and damage is visualised via thermographic imaging.The self-sensing capabilities of biocompatible CNT/Ultra High Molecular Weight Polyethylene (UHMWPE) (Reddy et al. 2018) were studied for the first time (Arif et al. 2018), where highly strain tolerant and sensitive strain sensors based on carbon nanostructures(CNS)polydimethylsiloxane (PDMS) nanocomposites were reported.An analysis of the fatigue life of quasi-isotropic Carbon Fibre Reinforced (CFRPs) laminates (Vavouliotis, Paipetis, and Kostopoulos 2011) was carried out by evaluating the electro-mechanical response.The effect of Multi-Walled Carbon Nanotubes (MWCNT) was also investigated.A polylactic acid (PLA)/MWCNT nanocomposite microstructure is illustrated in Figure 19, which is part of a conductive element in circuits.A high potential was reported for the damage and load detection (Böger et al. 2008) of Glass Fibre Reinforced Polymer(GFRP) structures via   The effect of print direction in FDM on the mechanical property on GO-ABS nanocomposite was demonstrated (Dul, Fambri, and Pegoretti 2016) for low graphene loading of 0.06% with a decreasing stiffness (6% for facedown and 15% for upright), improving fracture strength (3.5% for facedown and 10% for upright) and toughness (20% for facedown and 55% for upright) while also increasing the higher strain-to-failure (14% for facedown and 29% for upright) for a dog-bone shaped tensile test specimen.A simplified sketch of the route to multi-functionality is illustrated in Figure 20.aiding in achieving the desired end products of 3D-printed multifunctional composites.It also showcases how different functionalities can be incorporated at the structural level.It plays a significant role in the Industrial Revolution 4.0 in the future.
The influence of different FDM process parameters like build direction, layer height and infill percentage based on carbon fibre-reinforced PLA was studied (Kamaal et al. 2021).A multi-optimisation operation was performed using TOPSIS (Technique for Order Preferences by Similarity to Ideal Solution) to obtain the maximum strength, tensile or impact with the minimum material.
Obtaining an ideal reinforcement content amount would be critical in determining the force required for extrusion and reducing the equipment's possible wear.This is evident from the micrographs in Figure 21, showing the various PLA/CNT composite films with increasing concentrations of CNT.The presence of voids between stacked printed filaments is a significant deterrent in FDM products and can lead to fluctuation in the electrical conductivity values (Weeren et al. 1995).In FDM printing, the material needs to have its temperature raised above its glass transition temperature (T g ), and the determination of the T g value is the most vital parameter (Utela et al. 2008) in this regard.It can affect the morphology, and mechanical performance of the product and the alignment of fibres in 3DP is a significant area of interest in the reviewed literature.Other issues include poor polymer matrix and fibre adhesion, void formation, and printing of continuous fibre composites (Parandoush and Lin 2017).
One of the major features offered by 3DP manufacturing techniques is the possible customisation, which indicates that individual parts built using the same technology can be distinct from each other.Therefore, it poses a challenge to determine the ideal processing parameters for each part or production run.This is also evident from work on several AM techniques including cold spray AM (Menon, Aranas, and Saha 2022), which is a specialised DED manufacturing technique for coatings.Metal-based cold gas dynamic spray coating involves metallic powder fed as feedstock for deposition onto a substrate in a supersonic/transonic atmosphere.The 3D coatings are formed through self-consolidation without any limitation on the thickness and is ideal for complex geometry shapes.The influential process parameters affecting the quality of the deposition included the type of material, particle morphology and gap between the nozzle and the substrate surface.
All the findings from the previous studies mentioned in this section illustrate how complex the design and manufacturing of advanced functional parts can be and highlights some of the challenges to be overcome.These can provide guidelines on exploiting the full capability of 3DP techniques and tuning made possible by incorporating AI/ML techniques.The ideal set of parameters can produce high-quality parts in the first go, eliminating the need for trial runs and subsequent elimination of wastages and quality assurance tests.A reliable AI-powered predictive capability based on multi-scale modelling can play a vital role in determining the defect-free microstructure of advanced composite materials manufacturing by 3DP.The transition to a digital manufacturing scheme would not be complete without eliminating the need for manual intervention at different stages of the AM process.In the current scenario, this has not been fully realised and introducing ML techniques is a great leap toward this objective of automation.ML techniques offer great potential in minimising waste, enhancing the quality of printed parts and accelerating process parameter optimisation, especially when new materials are introduced.

In-situ and real-time monitoring for anomaly detection
The 3D fabrication process, in general, is prone to several anomalies, mainly owing to its complex number of process parameters and their interdependencies.It mainly depends on the human expertise of the operator to produce a quality part, which may otherwise be inconsistent and unreliable.In-situ monitoring methods are a need of the hour, and several developments, including advanced experimental equipment, enhanced simulation methods and robust computer vision capabilities, are being developed to pace up the growth.
Surface defects are one of the most influential factors affecting the quality of the fabricated part.An early, reliable (93.15% accuracy) and continuous monitoring system of the surface was introduced (Chen et al. 2021) for a DED process, which could prevent deterioration of the part quality.The novelty of the work came from deploying an in-situ point cloud based on ML, thereby effectively making the operation autonomous and requiring no human or sensor intervention.An in-house software platform based on multi-nodal architecture was developed to carry out the multiple, parallel sub-processes like ML feature extraction.A combination of unsupervised and supervised ML techniques was used to detect surface defects.The unsupervised clustering algorithm isolated potential defect regions in the point cloud and fed as input to the supervised classification algorithm, which made a detailed examination of the existence and classification of defects (if any) in each potential region of the point cloud.Also, a comparative study was made using eight different classification algorithms, including ANN, RF, SVM, K-Nearest Neighbours (KNN), etc. KNN was found to produce the most defect identification accuracy with a value of 93.15%.The proposed methodology was also proved effective after experimental validation of the results.
In another work (Lyu and Manoochehri 2021), a 3D laser scanner was added to a FDM machine setup to continuously monitor the FDM printing process, in order to improve the in-plane surface quality and geometric accuracy.The 3D laser scanner generated a point cloud dataset, providing layer information such as part height and in-plane surface depth.This dataset was then processed by a CNN model to extract features, detect, and further classify in-plane anomalies such as over-and under-extrusion, with an accuracy of 90.08%.The feedback was implemented as input to an online control system to adjust the parameters of the FDM process automatically.A CNN model was also utilised (Zhou et al. 2017) for the surface defect detection and classification for a SL process involving sheet metal parts.The SL process basically involves adhesive layers to join thin sheets of materials including polymers and metals with the aid of heat and pressure, which can be then cut into a 3D form.The CNN model developed processed the image of the defect sites through a series of defect segmentation, extraction and sliding detection methods.The proposed framework was able to achieve an accuracy of upto 97.02% and it was shown to be a function of the learning rate and inspection time.
In several studies (Jin, Zhang, and Gu 2019;Jin, Zhang, and Gu 2020;Goh, Hamzah, and Yeong 2022;Petsiuk and Pearce 2022), the ability of ML algorithms to detect underlying features and patterns was leveraged to detect multiple defects concurrently.A camera was fixed on a cantilever structure onto the extruder nozzle in a FDM process to monitor the printing area.CNN models were trained on real-time images to analyze intra and inter-planar defects (such as delamination).An accuracy of 98% was observed for intra and 91% for inter-plane defects, which underlines MLbased systems' efficiency.Multiple anomaly detection was also efficiently reported (Scime and Beuth 2018) using an unsupervised learning technique developed for the L-PBF process.The model held a filter bank to the image in any case of anomaly and recorded it as a dictionary based on the response of the filter clustering.The image is then mapped as fingerprints based on the dictionary similarity.In the defect detection process, the fingerprint of the new image fed is compared to the fingerprints in the database.This method recorded a classification accuracy of 95% for six and detection accuracy for seven cases (including anomaly-free) of anomalies.Similar techniques were used based on real-time camera monitoring and subsequent image processing by CNN algorithms (Jin, Zhang, and Gu 2020) for the automated detection of interlayer defects such as delamination and warping.The results indicated that the algorithm could classify the progress levels of delamination and detect and predict the onset of warping.The nozzle offset height was calibrated such that the model observed an accuracy of 97.8% and 91.0% on the validation and testing data, respectively.A strain gauge setup was used to track the warping tendency.The work promotes research towards automated pre-diagnosis of defects without requiring human intervention.
One real-time detection system, known as a threedimensional digital image correlation (3D-DIC) camera, was demonstrated (Holzmond and Li 2017;Zhang, Liu, and Shin 2019) to have efficient application even in porosity defect diagnosis for L-PBF products.As illustrated in Figure 22, the part is continuously monitored during printing with real-time image streaming.The camera captures the geometry of the printed part and compares it with the computer model to detect in-situ errors.Several studies similarly focus on the in-process monitoring of data from the melt pool.After taking the melt pool features as input, a CNN was applied (Zhang, Liu, and Shin 2019) to predict porosity in a laser-based 3DP process.The advantage of CNN to process high-dimensional data from visual images of the melt-pool was utilised to obtain porosity as an output.The accuracy of the results was verified by applying the developed model to specimens of various porosity levels.Another study (Bugatti and Colosimo 2021) focused on the detection of hotspots caused due to thermal effects developed during the L-PBF fabrication using high-speed and real-time image monitoring systems.The outputs were processed using several ML techniques, including K-clustering, SVM and NN.A comparative study was then made based on factors such as computational cost, applicability, and sensitivity.The prospects of this study envisaged improving the fault detection speed and accuracy, as well as extending the applicability for other defects and acquisition equipment set-ups.
A similar camera-based acquisition technique was utilised (Ogunsanya et al. 2021) for the in-situ monitoring of droplet formation in the recently emerging 3DP technique known as Inkjet 3D printing (IJP), which comes under the category of binder jetting AM technique.Liquid colloid droplets deposited on substrates can be affected by various factors like fluid properties and process parameters.Therefore, a backpropagation neural network (BPNN) was developed, which considered drop features such as size, aspect ratio, presence, satellites and velocity and classified output droplet modes such as normal, no-droplet and satellite-based on computer vision.The BPNN algorithm framework was shown to classify with high accuracy of 90%, indicating the scope for improvement of process optimisation and predictive analysis for inkjet 3DP using digital twin models.
The acoustic emission (AE) technique is widely used for non-destructive testing (NDT) in various structural engineering applications.The extension of AE to detect and monitor the progression of flaws in 3DP is gaining acceptance.Detection of flaws at the fabrication stages offers several advantages, including less wastage of material and fewer chances of failure and damage.Data analysis and signal processing are the significant steps involved in the in-situ monitoring of the 3DP process.The integration of ML was shown using a CNN model (Hossain and Taheri 2021) to perform the image processing of the wavelet transformed spectrum of the AE data and classify the build conditions of a DEDbased 3DP process.A classification accuracy of 96% and validation accuracy of 95% was obtained for the CNN.Furthermore, the outputs were evaluated against SEM results, indicating that different material samples had varying physical conditions.In a similar study (Becker et al. 2020) on FDM machines, a specialised microphone was attached to the extruder to record audio data and subsequent processing by a long short-term memory (LSTM) predictive model.The feasibility of such a model to classify the audio events into six different categories was demonstrated, and a printing stage could be aborted at early detection of error.However, it was also indicated that the model had inconsistencies distinguishing between errors originating from incorrect nozzle height and fan noise.
An interesting approach was developed (Pandiyan et al. 2021) using a semi-supervised approach to overcome some of the challenges of the conventional supervised ML algorithms for defect detection, such as the need for the huge balanced datasets.A set of two generative CNN architectures, namely GAN and Variational Auto-Encoder (VAE) was utilised to process raw airborne acoustic signals captured from specialised sensors to detect any deviations from set standards.Abnormalities such as keyhole pores, balling and lack of fusion (LoF) pores were shown to be successfully detected in the L-PBF processing of Inconel 718 and differentiated from a trained defect-free state.The architecture based on VAE was highlighted to be marginally better (96% accuracy) than GAN (97% accuracy) in terms of defect detection, when factors such as computational resources, network size, training time required, and trainable parameters were taken into consideration.However, it is to be noted that there is a wider knowledge required in the process setups, and subsequent fine-tuning/optimisation of the architectures before generalising them across different L-PBF configurationsincluding powder size, composition, and process parameters.
The capability of the above framework was also even further enhanced (Drissi-Daoudi et al. 2022) to address the challenge of defect monitoring in multi-material L-PBF.A ML model was built from three materials-Stainless steel, Bronze, and Inconel, shown to be effective with a classification accuracy of more than 86%.It was also indicated that acoustic signals indicating the defect regions are material dependent, and cannot be considered as a generalisation of defects between alloys.However, the results from the CNN framework showed promise of classifying the type of alloy and defect concurrently with good accuracy.
An in-situ health monitoring and diagnosis framework for FDM was developed in a study (Nam et al. 2020) utilising multiple sensors: three accelerometers, three thermocouples and an acoustic emission sensor.The data-driven approach based on the SVM algorithm used data from these sensors to derive the root mean square (RMS) values, which were used to build the diagnosis models.Post-validation of the models showed that the non-linear SVM model utilising RMS values from the acceleration frame attained the best performance.In this comparative study between healthy and faulty FDM processes, the faulty state was considered by the uneven levelling of the bed attached to the build plate.This imbalance may typically cause twisted shifting and warpage in the 3D printed layers.The framework demonstrated its capability in assessing how the 3D printed specimens can be monitored and predicted in an industrial setup.Another study (Sendorek et al. 2020) focused on how continuous real-time monitoring can be used as an indicator for evaluating symptoms of failure.It focused on the operation of IoT devices in Industry 4.0 and how data collected from sensors (even those retrofitted in older machines) can be worked on to develop advanced algorithms for predicting failure.
From the various studies on several AM techniques, it can be deduced that AI/ML offers great potential for insitu monitoring in AM, owing to their object detection ability (Goh, Sing, and Yeong 2021).It is vital in flaw detection and quality control in fabricated parts.However, care is to be taken as it can be challenging for reasons such as the need for large datasets as training data and improving the reliability of models.This translates to developing advanced data acquisition and processing techniques, including sensors, highspeed cameras, and processing conditions.It also entails a deep knowledge of computer vision and the printing process, which may depend on the expertise of the human operator.This can pose a challenge to its practical implementation and may need appropriate strategic planning.

Prediction of part property
An advanced form of the recurrent neural network (RNN) termed as long short-term memory (LSTM) predictive model was employed (Zhang, Wang, and Gao 2019) as a deep learning technique for predicting the tensile strength property of a FDM product.During the layerwise fabrication process of FDM, the in-process signals were captured by multiple sensors, including an accelerometer and IR, and fed to the LSTM network as input.
Figure 23 provides a flowchart on how the sequential information from each layer was fed to each LSTM cell, and a forward path was used for interlayer communication.This data was processed and later combined with material properties and process parameters for property prediction of the final product.This type of methodologies has the potential to gain more insights into process-structure-property-performance relationships in 3DP (Wang et al. 2022), which can aid us in reducing vulnerabilities and improving reliability, quality, and safe adoption of 3DP in various fields of applications.
A Bayesian network classifier (BNC) was used to make a comparative reliability study of various designs for fabricating using micro-SLA.The structure to be printed was a metamaterial having inclusions with negative stiffness.Post-fabrication of each design specimen, individual variability was recorded experimentally (also modelled) for the material property (Young's modulus) and geometry of the inclusion.This combined information, termed manufacturing variability, was fed to the BNC to distinguish reliable designs (Morris et al. 2018).
A fatigue life prediction study (Bao et al. 2021) was made using a combination of synchrotron X-ray tomography and ML for metallic Ti-6Al-4V alloy produced using the selective laser melting (SLM), which is a proprietary name for the ASTM standard term L-PBF technique.It was one of the few works to focus on determining the combined effects of geometric features like defect size, morphology, and location, related to classifying a defect as critical or not, which could initiate a crack, affecting the fatigue life of the part.As summarised in Figure 24, the SVM model was trained using the geometric features obtained from high-cycle post-mortem fractography and advanced tomography, followed by defect-determined fatigue life prediction and lifetime verification.The paper also points out the need to include factors such as surface roughness and residual stresses to assess the structural integrity of 3D-printed metallic parts.
Another fatigue-life-based study (Zhan and Li 2021) on one of the typical metal materials used in aerospace (other than AlSi10Mg and Ti6Al4V), 3D printed stainless steel SS 316L, involved continuum damage mechanics (CDM) for its data-driven framework, including the effect of variation of L-PBF process parameters like laser scan speed, power, layer thickness, etc.The output database obtained from the CDM is used to train multiple ML models like ANN, RF and SVM, and the obtained results were validated against experimental findings.Further investigations were made to estimate the prediction deviations, inaccuracies, and fatigue lives with variations in the parameters of the ML models.Future directions indicated the need for consideration of heat treatment and microstructure in the analysis of 3D printed products.
One study (Gong et al. 2022) involving PBF of metallic Ti-6Al-4V parts focused on predicting the specific cutting power, which is part of the machining in the post-processing stage.It involved the utilisation of a combination of advanced ML tools, such as Linear Regression (LR) and eXtreme Gradient Boosting (XGBoost) approach and Principal Component Analysis (PCA) techniques.The microstructure evolution was studied through the grain morphology and feature extracted to form a structure-property linkage.Massive datasets involving data from Scanning Electron Microscope (SEM), Electron Backscatter Diffraction (EBSD), X-ray Diffraction (XRD) (captures residual stress) and cutting force codes were processed by the ML tools.The novel procedure developed attained a high accuracy above 99%, indicating the feasibility of predicting post-processing machining behaviour, especially in 3D printed metal parts.
The versatility of a process-property relationship model, based on a Bayesian learning approach, was demonstrated (Lu et al. 2022)   wire-arc AM (WAAM).The process is a variation of the DED technique having an electric arc as the heat source to melt the metallic wire.In the study mentioned above, an evaluation is carried out on the thermomechanical variables and process parameters to map out a hierarchical ranking based on relative influence.The framework was tested on three alloys-SS 316, IN 718, and 800H, using NN and RF algorithms (both providing an accuracy of 97%).The preheat temperature of the substrate was found to be the most vital among process variables, whereas the difference between the preheat and solidus temperature was found to be the most influential thermomechanical variable.
A novel methodology was similarly developed (Valente et al. 2020) to enhance the effectiveness and predictability of the cold-spray metal additive manufacturing technique by monitoring the 'flowability' of the metal powder feedstock.The cold spray additive manufacturing is a solid-state deposition process and necessitates sufficient flowability to prevent nozzle clogging and part waviness.The flowability is standardised in terms of Hall Flow rates and is a factor of various physical particle-level measurements like size, shape and distribution.These measurements were fed to train a Decision Tree (DT) model to estimate the flow rate.An accuracy of 98.04% was demonstrated using the MLbased developed framework termed Flowability On Demand (FOD) in classification according to flowability.The work presents a quick and on-demand classification of potential feedstock, which can aid significantly in the repair of critical components in various applications, ensuring quality control and cost savings.The framework developed also promotes materials research by faster investigation results even on the scale of a per particle basis.
Another interesting recent development (Mythreyi et al. 2021) is the ML-based prediction of corrosion behaviour of L-PBF fused Inconel 718 in both as-built and post-processed states.The secondary motivation of the work also included estimating the effect of each postprocessing treatment on the overall 3DP manufacturing cycle, which can lead to several advantages, including better cost and quality control, convenience, optimisation of process and promotion of wide-scale adoption of L-PBF.Corrosion is a vital metric in the quality assessment of a component, and the corrosion testing in the above work was distinct in that it was carried out in an electrochemical environment.Two kinds of corrosion tests were carried out: potentiodynamic polarisation (PD), which measures the electrochemical activity and electrochemical impedance spectroscopy (EIS), measuring the stability of the layer formed after exposure to the harsh environment.The test results from the above methods were utilised to build multiple ML models, including Decision Tree (DT), SVM, Precision/Recall (PR) and eXtreme Gradient Boosting (XGB) to predict the behaviour of the material in the specified environment.The most accurate model for each test was used to carry out feature importance analysis, mainly to determine the most influential parameters affecting the corrosion resistance of the fabricated Inconel 718.The study paves the way for better comprehension of underlying complexities in introducing corrosion-resistance materials vital for future applications.
An investigation into reducing defects in 3DP was made (Du, Mukherjee, and DebRoy 2021) by predicting the occurrence of balling, a common defect in L-PBF parts, using a combination of mechanistic modelling, experimental data and physics-informed ML.The mechanistic model provided the values of the vital variables affecting the physics of the defect formation, which were revealed by inspection of the experimental data.These values were fed to a physics-informed ML model that provided a hierarchical order of importance of the variables on defect formation.These ML results were shown to make a more compatible and real-time prediction with 90% accuracy of the defect before the experiment was conducted.The proposed framework has the potential to prevent other common defects such as lack of fusion, cracking, porosity, and other complexities, thereby leading to improvement of part quality.This is regarded as highly beneficial for application in fields such as biomedical, wherein the L-PBF manufacturing process of beta-titanium alloys (Swee Leong Sing 2022), involves various fluctuations making the prediction and process control a challenging task.
The capability of tuning mechanical characteristics of 3D printed parts by processing data from previous runs using advanced data sciences tools like neural networks and a stochastic optimisation method like genetic algorithm was shown (Goh et al. 2021) to be effective in the multi-material PolyJet printing technique.It is especially common to manufacture anatomical models with tissuemimicking features in the medical field.Given that the objective of such printing techniques is to obtain parts with maximum compatibility with the native properties of the targeted tissues/organs, the study (Goh et al. 2021) has achieved some success in forming a designmechanical properties relationship network to obtain parts with desired mechanical properties such as compressive modulus and shore hardness.The neural network was able to learn the effect of design parameters on the effective mechanical properties with a mean square error (MSE) of 0.98% for the value of compressive modulus.In contrast, the genetic algorithm searched the design space for the desired shore hardness.One limitation highlighted in the study was the limited sample sizes, which could be overcome by increasing the datasets to cover the whole design capability of the PolyJet process.
Although studies are being done on ML applications in deriving the structure-processing-property relations in 3D printed parts that need deeper insights and scalability, there has been notable progress through recent novel developments, which give incentive for future possibilities, including behaviour prediction, leading to higher efficiencies and better adaptability.

Towards intelligent recommender system and printability analysis
It is highly advantageous if a feasibility analysis to ensure printability is done before allocating resources for the actual 3DP fabrication process.The power of unsupervised clustering algorithms was utilised for a FDM process in a study (Ghiasian and Lewis 2020) to develop a novel recommender framework which can perform AM-readiness and design modification of AM component database.Initially, the unsupervised hierarchical clustering ML model classified the part candidates based on their similarities in the potential for fabrication by feature extraction and provided subsequent design modifications based on specific infeasible part clusters.It could reduce the geometric incompatibilities and make the transition to the production stage smoother and more efficient usage of resources.Several other studies have also focused on developing ML-assisted decision-making system to identify part candidates and predict printability in the conceptual phase.With the aim of making 3DP more suitable and appealing to non-experts, an automated decision support (DSS) system was developed (Yang et al. 2020) to determine the candidacy of a particular part design for 3DP.It was based on different regression ML models and specific candidacy criteria captured from databases or digital models.A comparative study was made with various regression models, and based on the RMSE value, the boosted decision tree (BDT) algorithm performed best for all the training data.
Along similar lines, Mycroft et al. (Mycroft et al. 2020) worked on printability by employing a MLbased framework to test the geometric limits of metal 3D printing processes like electron beampowder bed fusion (EB-PBF).Printability is a factor of several parameters, including material and quality of powder, geometry and process parameters.The study focused on the relation between geometry and printability by a data-driven approach.Initially, a detailed metric was devised for fabricating specimens outside and within the limits of printability, followed by an analysis to ensure compatibility.Finally, the geometric features and several ML techniques (SVM, RF, Autoencoders) were used to measure the printability of a given specimen.The limitation of the proposed predictive model was also indicated, citing the inherent stochasticity in a 3DP process.
The evaluation of printability is of vital importance, especially in metal AM techniques involving new alloys.The prediction of processing and printability map can be formed on the data extracted from sources such as experimental monitoring and inprocess signals.A well-defined printable area can lead to the efficient selection of process parameters to produce parts with the desired properties.

Challenges, research gaps and prospects
The above study pointed out the efficiency and potential achievable by introducing the current AI/ML techniques into the 3DP process workflow.However, there are several shortcomings, challenges, and research gaps, as illustrated in Figure 26, which need more attention and research to facilitate and adapt the data-driven AM approach across various domains.

Multi-scale modelling
The predictive multiscale design based on statistical ML techniques is considered to be a driving force for the development of next-generation advanced smart composite materials.It entails the unification of AI/ML, large-scale modelling and material informatics to develop a reliable workflow from design to production.However, this approach for 3D printed techniques entails various deficiencies and technical limitations, which are not addressed in many research studies.For one, the training data required for ML models are also obtained from high-fidelity simulations of the various processes in 3D printing.These simulations should be reliable and accurate, taking into account all influential factors at specific levels of resolution.
This parametric evaluation is a significant field of study as the current computing capabilities need to be improved to obtain solutions within acceptable timeframes.Software frameworks must be developed for each constituent 3D printing sub-processes and link the complexities of the various sub-models.Also, future works need to develop schemes to reduce and quantify the uncertainties associated with the developed models using validation techniques.There are various categories of advanced ML algorithms, and each of them may not yield identical results for cases of different material systems, processing conditions, etc.Therefore, as discussed in the previous sections, it becomes crucial that a profound knowledge is comprehended from the various research works using varied parameters.
Other shortcomings observed include: (1) Lack of input data with reasonable accuracy, for example, in cases of materials exposed to elevated temperatures (2) Inaccurate assumptions made to account for unknown influences and speed up computational efficiency (3) Lack of field or experimental data for validation purposes The above challenges could be overcome by evaluating the sensitivity of the process route to various parameter values such as laser power, scan rate, etc.It can guide in identifying the critical parameters from the trivial ones to improve process efficiency and accuracy.

Quality control
The primary aspect to be assessed in a 3DP process is the quality control of the fabricated part.Premature failure may occur due to voids, delamination and other sources which can adversely affect the intended functionality of the part component.Characterisation tests may not yield accurate values as expected.Several developments, including multiple sensors (thermal electrical), are inducted into the 3DP process as part of the quality monitoring and improvement.Therefore, more research into analysing the sensor information by ML techniques can prove significant for quality improvement.In addition, new simulation, experimental, and computer vision methods are continuously being developed to enhance the working methodology and effectiveness.As represented in Figure 27, a working methodology involves coordinating various fields, such as data science, physics-based domain knowledge and experimental data, to form an accurate and efficient surrogate model.
In one of the earliest works in automated vision inspection in 3DP systems, Wu et al. (Wu et al. 2016) worked on detecting infill defects which may be formed as a factor of the unique vulnerabilities of 3DP systems due to its capacity to affect the infill without influencing the external morphology.Image classification based on ML was employed to detect malicious infill defects.The source of images was the layer-bylayer capture of the software simulation preview from the top view.Data extracted from the images were fed to two ML algorithms: J48 Decision Tree (DT) and Naive Bayes Classifier, and J48 DT has a better accuracy of them with 95.51% for classification between defect and non-defect images.An advanced approach was demonstrated by Khan et al. (Khan et al. 2021), wherein an integrated camera with the 3D printer captured real-time images at specific intervals.A CNN model was used to capture the feature extraction of geometrical anomalies that may exist in infill patterns due to factors such as sagging, inconsistent extrusion, lack of support and weak infills.These characteristics were used for a comparative study with those of a perfect 3D printed specimen.This methodology, based on the techniques of image classification and computer vision employing ML, provides the pathway towards a more automated and optimised workflow for the 3DP process and addresses the critical concern of product variability in 3DP.
Several advanced researches have been carried out for integrating testing and fabrication into the 3DP process, including the automated testing and characterisation of AM (ATCAM) framework for the FDM process.(Mazhari et al. 2021).The study aimed at quality consistency and repeatability by considering factors like feed quality, accuracy and calibration of equipment which influence the 3DP process output.This sensitivity of the 3DP process provokes the characterisation of fabricated parts for an individual process sequence.Coupons of the 3D printed samples were impacted dynamically by the deployment of an in-situ actuator, which was fabricated on the build substrate.The ML comes into play in the impact sensing by the load cell set-up.The methodology was utilised to analyse the feed quality and fatigue of the actuator over different cycles for three different PLA materials, utilising three ML algorithms.The gradient booster regression (GBR) algorithm performed comparatively well with a 71% correlation, proving confidence in the ATCAM methodology.
A review (Wu and Chen 2018) conducted on the quality control approaches in 3DP highlights the issues to be focused on at various life cycle stages, including design, process planning, and quality control (QC) at various sub-steps of the printing process.
The three stages of QC included: . Incoming QC: Validating the suitability of incoming raw material .In-process QC: The role of specific signal detection during manufacturing which can be related to defects in finished products .Outgoing QC: The product is validated against the requirements of the customer It is imperative that the quality of 3D printed products be improved by such approaches, as it can lead to wider adoption and reduced investment in machinery, materials and production.However, other QC approaches, such as cause-and-effect analysis and Design of experiment (DOE), though popular industrially, were too subjective and may not yield the expected optimisation in all cases.Furthermore, QC techniques such as control charts and Taguchi's method (Yang and Basem 2008) require extensive experimentation, which limits their applicability.

Adaptive learning using uncertainty quantification
Uncertainties can inherently be present in a statistical technique, and monitoring them on a consistent timeline can improve the learning capability of the model and enhance its predictive nature.Therefore, a tradeoff scheme between 'exploration and exploitation' was introduced (He and Powell 2018;Lookman et al. 2017), where the model is initially designed to 'explore' the property space.Subsequently, the adaptive design facility then progresses the system to 'exploitation' once the predictions of the ML model improve and uncertainties are reduced.It was also pointed out how ML models quantify uncertainty at a point in both classification and regression tasks (Meng et al. 2020).In classification tasks, it is in the form of a 'confidence' term, whereas for models like Gaussian Processes (GPs) applied for regression tasks, the standard deviation can be a good indicator of the uncertainty at a particular point.Also, ML-based models can significantly aid in acquiring the massive number of datasets required for typical uncertainty quantification (UQ) procedures, which might not be practical to obtain from simulations and experiments.
An ICME framework based on CALPHAD (calculations of phase diagrams) was developed (Wang and Xiong 2020) for composition optimisation and uncertainty quantification in the metal alloy 3DP process.Prealloyed powder often diverges from the targeted composition, affecting the intended 3D printed product performance.The methodology involved high-throughput calculations to analyse the process-structure-property relationships for around 450,000 compositions around the nominal composition of the case study material (high-strength, low-alloy (HSLA-115) steel).The framework was demonstrated to improve the probability of reaching targets by 44.7% and underscores the need for predictive ICME models and quality CALPHAD databases.
In another interesting study, a mesoscale melt-pool model for a L-PBF AM process was demonstrated involving global sensitivity analysis, multi-fidelity modelling, intelligent design of experiments and discrepancy modelling (Ghosh et al. 2021).The focus was to quantify uncertainty, discover missing physics in the simulations and accelerate the manufacturing process parameter development using a probabilistic modelling approach which can efficiently combine simulation and experimental data.The framework involved tools developed by General Electric (GE) Research, such as GE Intelligent Design and Analysis of Computer Experiments (GE-IDACE) (Aggour et al. 2019), encompassing the intelligent sampling framework with optimisation and GE Bayesian Hybrid Modelling (GEBHM) (Aggour et al. 2019) as the probabilistic ML technique.
A sub-field of ML which is gaining rapid attention and research being done is active learning.Active learning aids in attaining maximum model performance with the least set of annotated data.For example, highdimensional data from sensor integration is often prevalent in the new advances being developed in 3DP.Although such data can be utilised by supervised ML to predict process quality, the amount of annotation involved requires labour and time investment.In such cases, active learning can be a very effective and efficient technique (van Houtum and Vlasea 2021).Furthermore, ML models can query interactively for new data labelling during training using this approach, unlike the conventional requirement of labelled data (Meng et al. 2020).This indicates the potential to improve performance even if the data points are low.
A novel active learning technique termed Adaptive Weighted Uncertainty Sampling (AWUS), as shown in Figure 28 was used for employing model change between iterations of active learning to balance random sampling with uncertainty.The main working advantage of this active learning technique is adapting from exploration of instance space to exploitation of the model knowledge.The AWUS framework was used in the study to analyse a DED process in predicting process quality based on feature extraction and classification.The results were promising compared to random sampling, showing a reduction in annotations by 20-70% in 90% of the conducted experiments.Another approach (Xia et al. 2020) had a Model Free Adaptive Iterative Control (MFAILC) algorithm to model the dynamic process involved in a DED technique known as WAAM.An adaptive neuro-fuzzy interference system (ANFIS) based model was used to conduct simulations, and the results indicated the potential for enhancing automation of WAAM and forming accuracy.

Gaining more knowledge from limited datasets
The choice of a particular ML algorithm may be a function of the available dataset.Some may require a considerable amount of training data to attain optimum efficiency, but this size of the dataset may not be available from all kinds of 3DP processes.It can affect the reliability and accuracy of the models developed using such sparse data.However, new developments of models such as GANs (Creswell et al. 2018) are showing comparatively better results in such cases.Also, GANs have been shown to be well-suited for inverse material design (Lee et al. 2021), where properties are enlisted and candidate materials are evaluated and identified.
GANs could be used to explore larger design spaces and locate ideal materials and molecules for specific applications (Tabor et al. 2018).Examples include the generative ML model (MatGAN) (Dan et al. 2020) based on GAN to generate new inorganic material designs.It was shown that hypothetical materials generated included entities outside the training set, attaining novelty values up to 92.53% for 2 million samples.Large-scale computational screening and expansion of design space for inverse design of inorganic materials were the potential applications of the developed algorithm.Studies involving deep generative learning are also being conducted for the inverse design of high-entropy refractory alloys (Debnath et al. 2021).It was shown to learn complex relationships, which can be used as a tool to produce novelty, substantiating its role in the field of material informatics.ML frameworks for inverse design are also being developed for the optimisation of lattice unit cells of lightweight metamaterials to achieve desired mechanical properties (Challapalli, Patel, and Li 2021).Also, ANN applies to cases with insufficient governing relationships (Fakhrabadi et al. 2011).It is computationally inexpensive and can be user-friendly for relatively novice users (Matos, Pinho, and Tagarielli 2019).
Another field where ML techniques are utilised to overcome the lack of data is bioprinting (Yu and Jiang 2020), a specialised 3D printing technique involving biomaterials, cells and related growth factors developed for various biomedical applications.ML models based on CNNs and GANs (Yu and Jiang 2020) were studied to optimise raw material combinations and concentrations, including the prediction of properties.This can reduce the dependency on experimental testing and aid in developing novel bioprinting material and techniques faster.The potential benefits of maintaining well-curated Big Data and building Digital Twins are also indicated in recent studies, especially in specialised fields like 3D bioprinting (Goh, Sing, and Yeong 2021).New findings (Goh, Sing, and Yeong 2021) have underlined this by highlighting the major transition of 3D printing to the digital scheme in the future, where the maximum reliability is realised for the predictive power of the digital twin of a specimen, for example, a human organ, based on curated Big Data.It is to be noted that the right balance between the physical and virtual experiments would produce the maximum effective and efficient utilisation of resources.However, there needs to be corresponding development of infrastructures, computational power, and personnel training to accelerate these AI/ML applications.

Making more compatible models
A specific optimised model with a set of parameter settings and geometry may have a limited scope of application across different processes.It may have restrictions for different geometries, materials, and printing systems.This restricts its usage and causes additional repetitive work for the designers.ML can potentially eliminate the need for numerous computational simulations or trial-and-error experiments, which may conventionally be followed while transferring to a new printer.A data-informed starting point built on previous existing data can form the basis of a data-mining-based ML knowledge transfer framework.The model transfer learning approach by Sabbaghi & Huang (Sabbaghi and Huang 2018) was one of the earliest works to explore this domain to improve the comprehensive modelling of distinct processes.An equivalence framework based on the Bayesian method enabled the transfer of the deviation model (for geometrical accuracy) across different processes, even though the experimental data was limited.
The transferability of process-property relations was shown to be feasible (Lu et al. 2022) in L-PBF technique for two Inconel alloys.The process-property model of Inconel 718 was used to adapt for Inconel 625 using a Bayesian leaning approach, with a high accuracy of 0.95 R-value developed, indicating the feasibility of the procedure.Another distinct finding was how the amount of data on Inconel 625 affected the accuracy of the knowledge transfer model.An improvement in accuracy, i.e. a decrease in Mean Absolute error (MAE) and Root Mean Square Error (RMSE) on relative density recorded at 0.35% and 0.45%, corresponding to a 15-50 increase in data size.
A similar study (Liu et al. 2021) was directed toward ML-based knowledge transfer across multiple L-PBF printers of the same manufacturer using similar/ different technology and different manufacturer with different technology.These developments could save the huge cost and time involved in the extensive simulation and experimental procedures in optimisation while replacing existing printers or introducing new metal 3D printers.A global multiparametric optimisation algorithm was developed to simultaneously optimise multiple properties and demonstrate the capacity for cross-machine knowledge transfer.Parameter recommendations based on these models, which can operate on limited existing data to optimise properties by adjustment of model precision to the statistical significance of source data, can act as a starting point for more rigorous studies, such as sequential optimisation.In this study, the Bayesian model was found to be better than logistic regression and SVM in the case of printing Ti-6Al-4V after verification of property predictions.Prediction capabilities of ML-powered 3DP also need to be expanded for non-linear behaviours such as damage and crack evolution.
A machine-agnostic algorithm was successfully developed (Scime et al. 2020) for layer-wise detection and classification of an anomaly during various powder bed AM processes.It involved a novel CNN-based architecture for real-time pixel-wise semantic segmentation of layer-wise imaging data.
Advantages of the developed algorithm include: . Seamless transfer learning of knowledge between various AM machines and successfully demonstrated with techniques such EB-PBF, L-PBF and binder jetting .Accurate real-time performance .Ability to produce segmentation results at a similar resolution as the imaging sensor.

Enhancing data pre-processing and thermal image processing capabilities
Data processing is a vital process in systems based on ML.Data pre-processing maybe required on the input data, for example, in cases of 3DP where different printing parameters can have varying levels of influence on the intended output property.Feature selection and combination from the processed data by ML algorithm can prove to be beneficial.However, pre-processing data still has several challenges, especially in extracting information like a crack distribution from images obtained from experimental equipment like a scanning electron microscope.
As the evaluation of some 3D processes involves a melt pool, a vast amount of data is generated as part of the thermal imaging setup.The significant aspect of computational models for these processes involves the simulation of heat transfer and flow in and around the melt pool (Cook and Murphy 2020) as the powder bed progressively melts.One of the earliest studies to employ a physics-informed neural network (PINN) infusing first physical principles and data for temperature and melt pool dynamics prediction in the metal AM process was conducted (Zhu, Liu, and Yan 2021).The results indicated accurate predictions with a limited amount of labelled training datasets, proving its potential for broader adoption in advanced manufacturing.However, the study also highlighted some limitations in cases such as resolving evaporation, ambient gas phase and melt pool deformation of free-surface.Realtime control and in-process defect detection in PBF process using camera-based melt pool monitoring (MPM) is promising for ensuring the quality of built parts.Recent advances such as camera-based coaxial Microwave Power Module (MPM) for high-resolution data monitoring were evaluated for uncertainties, and effectiveness was demonstrated for a deep understanding of the PBF process (Lu et al. 2020).Deep learningbased process monitoring using a classification model for thermal images in selective laser melting (SLM)/L-PBF technique (Kwon et al. 2020) and DED (Li et al. 2020) were studied and proven to be successful, even achieving a classification failure rate under 1% for 13,200 images in some cases (Kwon et al. 2020).However, the modelling and process of the thermal images is still a daunting task, as it can involve the chance of variation in melt pool size and centre.New efforts are required to consider these inconsistencies and make the results reliable.

Lack of standardised regulations
Any mechanical test or characterisation has standards to ensure reliability and repeatability to the process and the consequent results.There are several standards ('ASTM') set for the 3DP techniques.However, these do not align with the needs and demands of the current research progress, especially for evaluating components and processes.The characterisation process standards are yet to be established for structures like hierarchically developed metamaterials and lattices.Several research efforts have indicated that ML-applied systems show good accuracy for evaluating 3D printed products, and the scope of complex geometries is improving with an increased research interest in recent studies.
Several pilot projects are being carried out by the American Society of Testing and Materials (ASTM) and International Organization for Standardization (ISO) to define qualification standards covering the needs of many industrial sectors (Moroni, Petrò, and Shao 2020).One of them, ISO/TC 261 (Pei 2020), discusses the standardisation of all the major AM processes, including test procedures, process chains (materials, data, processes, software and hardware, applications), quality parameters, health and safety, supply agreements, fundamentals, environment and vocabularies.In addition, efforts are being carried out to set ISO/ASTM standards to address the challenges, and current gaps for post-processing (Lee, Nagalingam, and Yeo 2021) of metallic 3D printed parts.For example, the industry sets high requirements for surface roughness which may not be precisely achievable with current state-ofthe-art metal AM techniques like PBF and DED.Therefore, it is essential that adequate ISO/ASTM standards are benchmarked so that surface finishing of complex AM geometries can be carried out in an efficient and admissible manner.A comprehensive evaluation of various published and under-development standards for qualification and certification of metal AM was compiled in a review (Chen et al. 2022), where the status of regulations developed by international communities and organisations was studied and challenges were addressed.The work also shed light on modern aspects such as digital qualification and certification using the aid of cyber-physical systems, blockchain and communication networks, as part of the outlook on the potential of industrial digital transformation.
There is a vast amount of data generated by different AM users, and a great opportunity arises to improve upon the process, which can be realised by ensuring that the data is accessible.A strict set of rules are to be followed in data sharing and interpretation.For example, online data repositories such as Gyrobotm, TeamGrenable, Bfessler and Profbink (Wendo et al. 2022), to name a few, store 3-D printable designs for prosthetics and can be selectively accessed for data extraction under a specific criterion.The features offered include clear distinguishing information into name, creator, year, particular instructions, signs of further development, etc.Another specialised online community is the e-NABLE platform ('E-NABLE Community') which also gives open access to 3D printable designs.The original creators may update the designs and models developed from time to time with improvements in process efficiency, which can be a technical starting point for any new developers in the field and gain insight from their findings.
Considering the aspect of data sharing, as illustrated in Figure 29, establishing a standardised database with supportive infrastructure and ease of accessibility is the need of the hour for several domains like materials science, geoscience, etc.However, only the field of computer science has been able to achieve some tangible advanced progress.Any improvement in ML techniques necessitates the existence of known datasets, which are accessible, well-curated and accepted by the community, forming the standardised benchmark for comparison.The National Institute of Standards and Technology (NIST) ('Additive Manufacturing Benchmark (AM-Bench).NIST') has developed one of the foremost such projects, providing process-structureproperty (PSP) data for AM benchmarking purposes.However, it is to be noted that there is still progress to be achieved towards managing and unifying data for easy accessibility, maximum utilisation of data and the reuse of AM data.

Mitigating cyber intrusions
One of the latest intuitive concerns to be addressed, especially given that the 3DP processes are data-intensive and dependent on computerisation, is its vulnerability to cyber-attacks.Cybersecurity can prove vital as 3DP is widely employed even for safety-critical systems.The inter-dependability of the physical product lifecycle with the virtual/digital twin chain makes the 3DP system a cyber-physical system (CPS) (Schleich et al. 2017).Belikovetsky et al. (2017) introduced a comprehensive chain of possible attacks along the 3DP sequence for the first time.It includes modifying the design file, which sabotages the manufactured part, thereby dismantling the cyber-physical system (CPS) using that part.Another objective was to evaluate the various attack chains and determine comparative difficulty.The practical was also demonstrated by showing the case of a propellor blade in hovercraft with a sabotaged design and the propellor breaking away in experimental flight.Attacks may also affect printing parameters through the filament and thermal (including nozzle temperature) manipulations (Rais, Li, and Ahmed 2021), and their consequent impact may not always be detectable without mechanical testing.
Concerns regarding unauthorised production and counterfeiting of original high-quality parts were raised in some findings.(Yanamandra et al. 2020).The study demonstrated the feasibility of reverse engineering for 3D-printed polymer-based fibre-reinforced composites using innovations in 3D scanning and imaging technology.
The imaging-based methodology based on a micro-computed tomography (CT) and scanning electron microscope (SEM) captured not only the geometry but also reconstructed the tool path of 3D printing.The reconstruction was performed based on the fibre orientation analysis, using a combination of multilayer RNN and LSTM on the microstructure.The framework achieved high dimensional accuracy for the reverse engineering model with a difference of 0.33%.A Binarized Statistical Image Feature (BSIF) (Yanamandra et al. 2020) algorithm was employed to convert images to a binary image format for compression without compromising essential features.This BSIF method was also used in a similar study (Chen, Yanamandra, and Gupta 2021) for defect detection in 3D printed fibre-reinforced composites using ANN architecture.The ML algorithm used large image datasets from tomography methods as training data.Any deviation in fibre direction compared to the predicted value was considered an issue of the defect in the microstructure.A mean square error of 0.001 was achieved using a refined CNN model in fibre orientation prediction.
One proposed strategy (Yang et al. 2022) to detect cyber intrusions was analyzing the audio and video signals captured during printing.For example, the online monitoring system may detect the shift in patterns of the audio signals based on a pre-set reference standard and reconstruction of the path of the extruder using the video signals.Another method (Moore et al. 2017;Liu, Kan, and Tian 2020) developed the technique of continuous monitoring based on the 3DP equipment actuators' current consumption and a comparative analytical study to detect anomalies during the process sequence.The scheme was demonstrated on a FDM process, by tracking the power supply to the filament extrusion motor and X/Y/Z-axis motors and detecting any possible modifications to them and the design file.
Novel methodologies to secure the G-code against intellectual property (IP) theft and unintended design modifications are also being developed, including block-chain-based data storage and practical asymmetry encryption approaches (Shi et al. 2021).It was shown that the proposed framework could detect malicious tampering and significantly eliminate unauthorised access to the G-code in case of a face shield fabrication which is vital in the COVID-19 pandemic times.'Physical hash' methodologies in the form of QR codes (Brandman et al. 2020) are also being developed, which link the digital data to the manufactured part.It does not depend on the state of the network for attack detection and ensures IP protection for the toolpath and process parameters while also permitting in-situ quality assurance.Other solutions (Gupta et al. 2020) devised towards mitigating these cyber risks include the non-traditional formulation of a supply chain model for the 3DP, which includes the printer, raw materials and design files.A holistic integration may also include the virtual supply chain; however, the model's hierarchy may depend on the application domain and level of security infrastructure available.The role of collaborative work required between designers, engineers, scientists and security professionals is emphasised and can play a pivotal role in safeguarding the public and national security.

Conclusions
The paper has systematically reviewed the existing datadriven AM techniques encompassing various sub-stages and areas of interest along the process workflow.A detailed analysis along the process-structure-property chain has been made, entailing in-depth information on existing advancements and progress achieved.The study focuses on the future generation of AI/MLpowered models for smart composite material development, including multiscale simulation framework and manufacturing using advanced 3D printing techniques.The contemporary ML applications discussed in the paper included the multiscale design for additive manufacturing (DfAM), parameter optimisation, in-situ monitoring in the area of interest during the process, part property prediction, quality assessment and control.The existing shortcomings in the various stages of several state-of-the-art 3D printing techniques have been identified.It includes technical and non-technical challenges such as quality control, uncertainties, model compatibility, image processing capabilities, lack of standardised regulations and cyber intrusion.The potential solutions discussed contribute toward the standardisation of ML-integrated 3D printing techniques.Each of the discussed methods involves evaluating the process route to determine the process sensitivities and related critical parameters for enhancing the process efficiency by using targeted experiments/simulations.
The application of several AI/ML techniques is discussed at each stage of the process chain in several AM processes.Each stage of the process generates a vast amount of data which can be utilised to train ML algorithms for intended results.A holistic integration involving the application of ML at various stages of the AM process can lead to better consistency and quality control in 3D printed parts, thereby improving their reliability and potential adoption in industrial applications.The study has highlighted that new advances in data acquisition, processing, and data mining from existing information are effective in exploring new material designs and product development.The whole process necessitates the collaboration of various teams performing data analysis and prioritisation at different stages of the design and manufacturing process.This coordinated knowledge-sharing approach promotes collaboration within the research community, making knowledge more accessible, interpretable, and open to innovations.
The current challenges in potential applications of data-driven AM have also been identified, and efforts should be mainly focused on improving the quality of large datasets-including developing advanced ML algorithms with precise accuracy and higher computational power.The very prospect of establishing a reliable AI-guided production workflow, from the Representative Volume Modelling (RVE) to the 3D printing fabrication stage, offers immense potential.It should also be considered that any discrepancy in the predictive outputs of ML models can adversely affect the whole process chain of any digital manufacturing such as AM.Therefore, ensuring factors such as transferability, data-driven and robustness in the developed ML algorithms form the critical factors in eliminating errors and ensuring the effectiveness of ML-integrated additive manufacturing.
Future research should emphasise improving the reliability of predictive models to fully realise the design freedom offered by AM with maximum efficiency.The various areas of interest discussed in the paper, including the concept of active learning, uncertainty quantification (UQ) and improving AM data, can accelerate the integration of ML techniques in AM and pave the way for future opportunities in high-end applications of 3D printed parts.

Notes on contributors
Sandeep Suresh Babu is a PhD student and Research assistant at United Arab Emirates University.He obtained his Master degree in Aerospace Engineering from Indian Institute of Technology Bombay (IITB), Mumbai, India and Bachelor degree in Aeronautical Engineering from Anna University, Tamilnadu, India.His research interests span across multiscale modelling, structural modelling, finite element analysis, 3D printing, Bioprinting, shape memory materials, composite laminates, metamaterials, biomaterials, material characterization, and machine learning for material development and manufacturing.His research has resulted in more than 350 referred and peer-reviewed international journals and conference papers and book chapters.His current research interests include nanomaterials and nanotechnology, photovoltaics, supercapacitors, additive manufacturing, materials characterization, tissue engineering and biomaterials, stress corrosion cracking, fatigue and fracture of materials, stress analysis, finite element method, failure analysis, and prevention, durability and degradation of polymeric and composite materials, metal matrix composites, friction stir welding FSW and friction stir spot welding FSSW and hydraulic expansion joints.He has been also honored as a 'Distinguished Professor' at UAE University in 2022.

Figure 2 .
Figure 2. A simplified ML workflow: Operation to draw insights from data.

Figure 4 .
Figure 4. Research direction related to this study.

Figure 6 .
Figure6.A classification scheme of machine learning algorithms based on their operating principle.A generalised application area for these ML algorithms is also identified for the various AM processes based on the past literature reviewed in this study.

Figure 8 .
Figure8.Scope of ML algorithm application: Various stages and parameters of the 3DP process(Jin et al. 2020).

Figure 9 .
Figure9.A typical workflow of a ML application in a material design problem(Guo et al. 2021).

Figure 10 .
Figure10.Scope of ML application at various stages of the additive manufacturing workflow.

.
Substituting numerical simulation by ML techniques: Even though both numerical methods and ML are based on mathematical modelling(von Rueden et al. 2020) and data analysis to predict system behaviour, ML is often advantageous because ML can model phenomena of higher dimensions at a faster rate.. Integrating simulation data into ML training framework: This may be done at different components of the ML framework, including algorithm, training data, hypothesis, etc.Similar to data from experiments, simulation can be used to provide training data for the ML model.For instance, simulation data was used(Rahman et al. 2021) to train the ML model to predict the shear strength of carbon nanotube (CNT)-reinforced composites and the variation was assessed.

Figure 19 .
Figure 19.Polylactic acid PLA/MWCNT nanocomposite fabricated via Liquid Deposition Modelling (LDM) (A) Optical micrograph of woven-like microstructure (Top view) (B) Top and (C) Side view SEM images of a 10-layer scaffold; (D) Optical photograph of LDM fabricated filament in freeform with a 1 cent coin for comparison; (E) 3D printed PLA/MWCNT nanocomposite woven structure used as a conductive element in a simple electrical circuit (Postiglione et al. 2015).

Figure 21 .
Figure 21.Polylactic acid-carbon nanotube (PLA-CNT) composite films: Optical micrographs of hot-pressed composite films fabricated by FDM process, showing the difference in texture with the increase in the concentration of CNT content (Patanwala et al. 2018).
using a knowledge transfer learning concept between Inconel alloys manufactured by L-PBF.The process-property relationship of a lesser-studied Inconel alloy, IN625, was captured using the knowledge obtained from another relatively better-known Inconel alloy, IN718.The model developed exhibited an exceptional accuracy of 0.95 R-value, thereby proving the robustness of the technique and the capability of knowledge transfer across materials, which can produce unprecedented impacts.Wu et al. (2020) utilised a unique combination of a thermomechanical model and multiple ML models, as shown in Figure 25, to mitigate residual stress development in a

Figure 25 .
Figure 25.Schematic flowchart of the residual stress evaluation process in wire-arc AM (Wu et al. 2020) where E and β represent Young's Modulus and coefficient of thermal expansion, respectively.

Figure 26 .
Figure 26.Schematic of the research challenges in adopting AI/ML techniques effectively into 3DP and future research directions.

Figure 29 .
Figure 29.The concept of data sharing and resharing.
Abdel-Hamid I. Mourad is currently a professor at Mechanical Engineering Department, United Arab Emirates University, UAE.He obtained his Bachelor degree and M.Sc.from Helwan University, Cairo, Egypt.He received his Ph.D. in Mechanical Engineering from Indian Institute of Technology, Bombay (IITB) in 1995.He has been awarded 'Distinguished Professor' by UAEU on 2022.He is recognized as TOP 2% scientist by Stanford university.He is a member of Member of Mohammed bin Rashed Academy of Scientists (MBRAS), UAE.
Khalifa H. Harib is a professor of mechanical engineering at the UAE University which he joined in 1997.He also taught in the-Mechatronics master program at the American University of Sharjah as an adjunct professor in 2010.He earned his Master and PhD degrees from the Ohio State University, both in mechanical engineering, in 1993 and 1997 respectively.His research and teaching interests include robotics, computer aided manufacturing, dynamics and control.Prof. Harib has authored and co-authored numerous journal and conference papers and book chapters.He chaired the department of mechanical engineering at the UAE University from 2000 till 2003, and has served on various national committees including Dubai Metro committee and the ICT fund committee.Sanjairaj Vijayavenkataraman is an Assistant Professor of Mechanical Engineering and Bioengineering at New York University Abu Dhabi.He is also affiliated with the Department of Mechanical and Aerospace Engineering at Tandon School of Engineering, New York University, Brooklyn, USA.He is the founder and director of The Vijay Lab at New York University Abu Dhabi, with a focus on 3D printing and Bioprinting for tissue engineering, regenerative medicine, drug testing, and medical devices.He received his PhD from the Department of Mechanical Engineering at National University of Singapore (NUS) in 2019.He was also part of several life sciences and biomedical industry-oriented programs such as Singapore Stanford Biodesign (SSB) Innovation Class, NUS Lean Launch Pad Singapore (modelled after NSF I-Corps program), and P&G Serial Innovator Camp.He was the recipient of several prestigious fellowships and awards including the President's Graduate Fellowship, which is awarded to candidates who show exceptional promise or accomplishment in research, for his doctoral studies at NUS (Singapore), DAAD WISE Scholarship from Germany, Raman Memorial Award and The Sachivothama Sir C.P.Ramasamy Aiyar Scholarship during his undergraduate years.His research interests include Additive Manufacturing, 3D Bioprinting, Electrohydrodynamic Jetting (EHD-jetting), Biomaterials (Polymers, Ceramics and Metals) for Additive Manufacturing, Tissue Engineering and Regenerative Medicine, 3D Printed Scaffolds and Bioprinted tissue constructs for Tissue Engineering, Regenerative Medicine and Drug Testing.

Table 1 .
A classification of various 3D printing process techniques based on their working principle.

Table 2 .
Application sub-fields of various ML-integrated 3DP techniques.