Effect of process parameters on the strength of ABS based FDM prototypes: novel machine learning based hybrid optimization technique

Even though the prototypes built using Fused Deposition Modelling (FDM) process are found to exhibit good mechanical properties, there are ample scopes to improve them by means of selecting suitable process parameters. Since the FDM process involves more number of process parameters, the selection of optimized values becomes more complex and time consuming. Further, the complex correlation among the process parameters makes the selection process more tedious and involves more numerical steps. Hence it has been intended to perform a physical experiment with the known parameters to determine the performance measures of the built prototypes. With this moto, in this work the effect of the 3D printing parameters is studied and the optimal combination of these parameters are determined. The Taguchi L18 orthogonal array based values are assigned for process parameters and the physical prototypes are fabricated. These specimens are tested in the laboratory and the observations are analyzed. It has been found that the process parameters under consideration have a good effect on the strength of the built models. Out of the 18 experiments, better experiments are selected by using a Machine Learning (ML) approach namely decision tree (DT). Finally, the best combination of parameters has been determined by using a novel hybrid multi objective technique which is formulated by integrating Fuzzy Analytical Hierarchy Process (FAHP) and Complex Proportional Assessment of alternatives (COPRAS) techniques. Then a confirmation experiment has also been done to confirm the optimal combination of parameters. The influence of the parameters is also found by using ANOVA (Analysis of Variance) method. The final results show that the raster angle influences the outputs more while the raster to raster gap has the least influence.


Introduction
The evolution of 3D Printing (3DP) technologies has changed the production chain procedure from the conventional stages. Custom products with complex geometries can be directly built by applying computer aided design (CAD) model data. This leads to less inventory cost and changes the paradigm from mass production to customized production. Hence these 3DP technologies have become the key feature in the entire value chain of the products in the Industry 4.0 [1]. The process of 3D Printing has been developed from the concepts of Topographic Modeling and Photo sculpture. Earlier, the term 3D Printing was classified as a part of Additive Manufacturing (AM). But due to the developments and inventions that happened in the field of three dimensional printing technologies, the term 3DP has been considered as an interchangeable technology with AM. Among these technologies, the Fused Deposition Modeling (FDM) process invented by Scott Crump in the year 1987 offers unique advantages such as less waste production, limitless customization, need for less skilled labour, and requires minimal capital investment when compared to other conventional processes. The FDM process uses coiled material (filament) and a moveable head with heated nozzle(s) to build the material in a layer by layer manner on a moveable build platform. 3DP processes are rarely used to build end components and are mostly used to build prototypes. 3DP exhibits the possibility of producing very complex prototypes with or without different material compositions [2]. FDM process uses materials like ABS (Acrylonitrile Butadiene Styrene), PC (Polycarbonate), PPSF (Polyphenylsulfone), PC-ABS blends, and medical grade PC blends. These materials are found to be the most cost effective. This process is a slow process when finishing is given importance. 3DP processes have the following capabilities as advantageous when compared to the conventional machining: Design flexibility, less geometric cost, single part assembly with relative motion, lower inventory, higher productivity, structure feasibility, leads to customized production, remote manufacturing and IoT enabled [3]. The co-polymeric structure of ABS material provides good mechanical strength, hydrophobicity, and chemical inertness to the build prototypes. Further due to the amorphous phase of styrene-acrylonitrile the ABS is found to be more stable to cracking because of stress, structural damages due to thermal loads, and chemical effects. Due to the polybutadiene phase, ABS is found to be good at toughness [4].
Dewada and Telang [5] have reviewed the advances of nanocomposites used in the FDM process and compared the performance of the polymer nanocomposites with nanofillers in view of mechanical properties of the build object. Swetham et al [6] have used a composite termed as 'carbomorph' and proved its significance in producing electronics embedded sensing devices in a single build process. Khoo et al [7] have insisted on the importance of more research works in overcoming various challenges such as the usage of smart materials and structures and they have discussed smart nanocomposites, shape memory alloys, and shape memory polymers. They have also elaborated the scope of new research works and future trends in 4D printing, especially in 4D bioprinting. Wong and Hernandez [8] have reviewed the classification of additive manufacturing and its applications in various fields and dealt with the concepts of different AM processes as well as their related applications in the real time production environment. It has been focused that there are lots of scopes for doing researches in selecting suitable build material, the level of accuracy needed to eliminate the finishing process, and the optimization of process parameters to yield better strength for the built prototypes. They have concluded that the AM processes would occupy a significant place in the future manufacturing scenarios. Tang et al [9] have experimented the effect of raster angle, rate of deposition of material, layer thickness, orientation, and diameter of the nozzle on the built 3D prototypes. The output parameters determined in the study were the tensile strength, elongation, shear strength and flexural strength. Materials such as ABS, PLA, PEEK resins, Z-UltaT, and HIPS were added and it was concluded that there is a lot of scope for pursuing further research work in optimizing the 3D printing process parameters concerning the mechanical properties of the built 3D components. Kozior et al [10] have studied the surface finish and mechanical strength of FDM samples made by 3D printing technology. First, the samples have been prepared by fused deposition modelling technology. The material used is polylactic acid (PLA). The samples were exposed to heat and chemical treatment through the use of acetone. The surface quality and mechanical properties were studied before and after the treatments.
Rismalia et al [11] have conducted experiments using PLA material with two input parameters such as infill pattern and density corresponding to the tensile property as the output performance, and found that the highest tensile properties were obtained for the concentric infill pattern. Banjanin et al [12] have conducted experiments to study the consistency in the mechanical behaviour of FDM fabricated components using PLA and ABS materials. The ABS materials based components are found to be consistent in their tensile properties while the PLA materials based components are found to be consistent in compressive properties. Arrieta et al [13] have used two different approaches such as global indexes and local surface distribution of errors to determine the geometric accuracy level of RP models. They have further proposed that there are avenues for research in reducing the processing time, topology control of the slices, and enhancing the process of segmentation. Christiyan et al [14] have determined tensile and flexural strengths of 3D prototypes built using a composite that constitutes ABS material and hydrous magnesium silicate as par with ASTM D638 and ASTM D 760 standards. From their experimental outcomes, it was concluded that low printing speed and lower layer thickness have been found to be more dominant than other process parameters and they have resulted in maximum tensile and flexural strength on printed prototypes. Padzi et al [15] have performed experiments using two types of specimens. The dog bone shaped part has shown lower fatigue cycles compared to moulded parts and they it was concluded that the tensile property of 3D printed parts could be enhanced by properly selecting the optimal process parameter. Tharun kumar et al [16] have explained the effect of FDM process parameters such as build orientation angle, model interior, and direction of rotation on the impact strength of built models and found that the angle of orientation shows the major effect on impact strength. Panda et al [17], have opined that full scale application of additive manufacturing processes is lagging because of the available materials for the rapid prototyping process. This issue can be solved by using newly developed materials in the RP process and by modifying the input process parameters during the fabrication process.
Taguchi method has been used by several researchers to optimize the welding parameters [18,19], and 3D printing [20]. Patel and Kadia [21] have adopted the Taguchi experiment and ANOVA to determine the influence of process parameters like raster width, contour width, layer thickness, air gap, and raster orientation on tensile, flexural & impact strengths and hardness of the prototypes built using FDM process. Tontowi et al [22] have stated that the default settings of the process parameters lead to dimensional error and reduce the strength of the printed parts. Three process parameters i.e., layer thickness, temperature, and the raster angles were optimized by using Taguchi and Response Surface Methods of components printed using PLA (Poly Lactic Acid) material. The results have shown that the layer thickness has a direct influence on the tensile strength and the raster angle values have a greater impact on the dimension errors. Jaya Christiyan et al [23] have used the composite consisting of ABS and hydrous magnesium silicate materials to build the prototypes as par with ASTM D638 and ASTM D 760 standards with various layer thicknesses and printing speeds. It is proved that the printing speed has more influence on the mechanical properties of built components. Wu [24] has proved a strong correlation between layer height with printing time, and precision in the FDP. It is concluded that 0.14 mm of the layer height resulted lower processing time and the print quality was ensured.
Rayegani and Onwubolu [25] have obtained the correlation between the process parameters and the tensile strength and found that the forecasted values are closely nearer to the measured values. Further, they have used the Differential Evolution (DE) method has been used to achieve good strength for the response of the mathematical model that they have framed and concluded that the optimization of process parameters plays a major role in building quality printed parts. They have also used FDM ABS (Acrylonitrile Butadiene Styrene) material for fabricating the prototypes using the Fortus 400mc System. Vishwas. and Basavaraj [26] have analyzed the relationship of process features with the ultimate tensile strength and the dimensional accuracy of specimens built using ABS material in the FDM PRAMAAN Mini machine. They have considered model orientation, layer thickness, and shell thickness as process parameters and used Taguchi's L9 orthogonal array to conduct experiments. They have shown the impact of orientation angle and thickness of shell on ultimate tensile strength and the dimensional accuracy with manufacturing time. In the FDM process, the wire type built material is allowed to pass through a heated extrusion header which is maintained at 270 degrees centigrade. While passing through the header, the material gets heated and reaches the molten state and this viscous material is allowed to deposit on a platform which can be lowered after scanning each layer. The platform acts as a fixtureless foundation for manufacturing the prototypes [27]. The major advantages of this kind of system are precision, speed, and reliability. With the fast growing applications of artificial intelligence and ML methods in materials science applications, the materials informatics are entering into the data-driven era [28]. Gupta et al [29] have developed a predictive model for the density of hybrid nanofluids using decision tree technique. The results have proved that the ML technique has predicted well than the conventional methods.
From the above comprehensive review, it is perceived that only a few works have been reported for the selection of process parameters in the FDM process. As the need for reliable mechanical properties of the built components plays a vital role in the current industrial scenario, more emphasis has been given to select optimal printing parameter values of the FDM process. Five numbers of input process parameters such as thickness of the layer, part build orientation, raster angle, width of the raster, and raster to raster gap are considered for optimization with respect to tensile, flexural and impact, strengths. A theoretical study has been conducted by considering five process parameters with three levels for each parameter using Taguchi Technique by constructing an L-18 orthogonal array. The same study is experimented with real time fabrication of the specimens and tested in physical environments by maintaining ASTM standards. Finally, a hybrid multi objective model is developed to find the better combination of parameters to yield optimal mechanical properties.

Materials and parameters
Acrylonitrile Butadiene Styrene (ABS), a polymer made of acrylonitrile, butadiene, and styrene and it finds a predominant position in prototyping applications is the material used in this work. It has better resistance to chemicals, heat stability, impact strength, toughness, rigidity, and printing properties. The ordinary material is in opaque ivory colour. ABS is 100% recyclable and is non-toxic as well as harmless [14][15][16]. The material is processed in Raise3D N2 series 3D printer which has a build volume of 305 × 305 × 305 cubic mm. The filament size of the printer is 1.75 mm and the maximum temperature of the extruder is 300°C. The parameters governing the mechanical properties are layer thickness (the thickness of the molten material deposited during each pass or the difference between two adjacent layers in mm), part build orientation (the orientation of the prototype concerning the build table in degrees), raster angle (the orientation of raster related to the longitudinal axis of the build platform in degrees), raster width (the pattern of raster adopted to fill the inside areas of part contours in mm) and raster to raster gap (the air gap between two consecutive rasters on the same layer in mm) [23,24]. The span of these parameters are given in table 1.

Taquchi's experimental design
The experimental works are designed by using Taguchi's concept in order to reduce time and cost. Since five parameters with 3 levels are considered, L18 orthogonal array has been selected for the experimental works [30]. Table 2 depicts the combinations of parameters in the L18 orthogonal array. The material is processed in the 3D printer with the combinations shown in table 2 and the tensile, flexural and impact strengths of the material are tested. The testing equipments and the environment are shown in table 3.

Novel hybrid multi objective optimization technique
A novel hybrid multi objective optimization technique by integrating DT, FAHP, and COPRAS has been developed and used to find the optimal combination of parameters. The research outline followed in this work is shown in figure 1.
All the needed data are collected through the 18 sets of experiments defined by L18 OA and the optimal combination is obtained in three stages as follows: (i) Selection of better experiments using DT.
(ii)Computation of the weights of the output performances using FAHP.
(iii) Determination of the optimal combination of parameters using COPRAS.

Decision tree (DT)
A decision tree is one of the ML techniques that is applied to make a decision in multi objective decision making environments. It is a tree-shaped classifier classified under the supervised learning approach. DT is implementd in both classification and regression problems. In the DT structure, the charecteristics of a dataset are denoted by the internal nodes, the decision rules are represented by the branches and the outcome is denoted by the leaf node. Decision node and leaf node are the nodes available in DT. The decisions or conclusions are made by using decision nodes whereas the ootcome of the decisions are represented by leaf nodes. The charecteristics of the given dataset are influencing the decisions made. The possible solution (decision) depends on the conditions considered is depicted graphically in DT. It is obtained by a Yes or No response for a decision question, and the tree is divided into subtrees. The procedure of DT is as follows:  Initially, the basic classes for the decision are derived. The entropy is calculated for all the classes using the Yes or No responses by using table 4 and the root node is obtained.

• Calculation of Entropy for the classes
The impurity or unpredictability in a set of investigations is measured by using the Entropy. It is used to split data in a DT [31]. The entropy for the output (E OP ) is determined by using equation (1).

• Calculation of entropy for other features after split
The entropy after each of the split for the other features (A 1 , A 2 , K.. A n ) is determined using equation (2).
where OP = Output; C i = ith feature / criterion; P C i.yes ( ) = probability of ith feature/criterion for condition 'yes or agree'; E C i.yes ( ) = entropy of ith feature/criterion for condition 'yes or agree'; P C i.No ( )= probability of ith feature/criterion for condition 'no or disagree'and E C i.No ( )= entropy of ith criterion for condition 'no or disagree'.

• Information gain for each split
Then, the information gain is determined for each of the features using equation (3).

• Perform the splits
After that, the feature with the highest information gain is put in the first split of the decision tree and the other splits of the tree are constructed using the similar concept till the final conclusion / decision is made. This is achieved by dividing the original table into sub-tables.

FAHP
FAHP is a combined technique in which the analytical hierarchy process (AHP) [32] and fuzzy concepts [33] are integrated [34]. FAHP is an effective methodology to find the criteria (output responses) weights. First, the output responses are compared together based on the perception of the experts. These comparisons are called pairwise comparison matrix/criteria matrix [equation (4)]. To quantify the comparisons, Satty's [32] nine-point scale is used (table 5).
where, x ij denotes the comparative importance based on the comparison of ith and jth responses & 'm' denotes the number of output responses. The fuzzy criteria matrix is formulated from the criteria matrix using equivalent triangular fuzzy numbers (TFN) [34] and they are shown in table 5.
Then the fuzzy criteria matrix is preprocessed/normalized by using equation (5). This is performed to bring all the data into a small range.
The weights of the output responses are calculated by the row average of the particular output responses in the normalized matrix. Since the criteria matrix is formulated from the expert's judgment, it is a must to validate the consistency. Equation (6) is used to determine the Consistency Ratio (CR) which is used to validate the consistency of the developed model.
In which CI is Consistency Index and that is determined using equation (7) and RI reveals the random indices for output responses 'm'.

CI m m
where max l is the maximum Eigen value. RI is estimated by Saaty [32] and that is shown in table 6. If the CR is less than 0.10, then the decision maker's pairwise comparison matrix can be accepted.

COPRAS method
COPRAS (Complex Proportional Assessment of alternatives) is a multi objective optimization technique to solve problems with similar as well as conflict objectives [35]. The procedural steps of COPRAS technique are presented below: • Formulation of decision matrix (DM) The experimental results (output responses) are presented in matrix form [equation (8)] which is called as decision matrix (DM).

DM OP OP
.OP OP OP .OP : OP : OP : OP 8 11 12 1m where n denotes the number of experiments.
• Normalization of decision matrix DM ( ) Handling of the raw data is very difficult. In order to convert the range of the raw data into a uniform scale, the raw data are per-processed/normalized using equation (9).  • Determination of the weight of the output responses (W j ) The major part of the multi objective optimization problems is the determination of weights. FAHP is applied to compute the weights in this work.

• Computation of maximizing index (Pj) and minimizing index (Rj)
Depending upon the qualitative nature of the expected outcome of the responses, the maximizing index (P j ) and the minimizing index (R j ) are determined. That means, for the responses with the maximization objectives (maximum is the optimum), the maximizing index (P j ) is found out and the minimizing index (R j ) is calculated for the responses for which the minimum value is the optimum. P j and R j are determined by using the equations (11) and (12) respectively. Finally, the ranking is done from the highest Q j [equation (13)] to the lowest value. The experiment with the highest Q j (i.e. Ranked as 1) is selected as better.

Results and discussions
The experiments have been conducted as per the L18 combinations and the observations are tabulated in table 7. It is also called as decision matrix [equation (8)     impact on the tensile strength of the models. The raster width and the raster to raster gap have a medium impact on building stronger prototypes. It is found that all the parameters have a good direct impact on the flexural strength of the built prototypes. The low values of Layer thickness produce prototypes with the least impact strength.
The part built orientation has a medium impact on building stronger prototypes. The values of raster angle and the raster width have a direct proportional impact on the impact strength of the models. The parameter raster to raster has the least significant effect on building models with good impact strength.
Since the Taguchi technique is a single objective optimization method, it is not possible to make an optimal decision. To rectify this issue, a multi objective optimization method is used. In this work, a novel hybrid method has been proposed to find a better combination of parameters. The weights of the output responses are calculated by using FAHP as mentioned in section 2.3.2. The fuzzy criteria matrix is presented in tables 8 and 9 depicts the weights. The consistency ratio [equation (6)] is determined as 0.06041 and the results are validated. Then the best experiments out of the 18 experiments are identified using the decision tree technique (figure 5) as discussed in section 2.3.1. The DT is constructed using table 7. First, the average of all the output responses are calculated. Then all the values are categorized into high (value > average) and low (value < avearge) to form the initial table(table 10).
From the information received from table 10, the decision tree is constructed as shown in figure 6. From figure 6, the experiments 6, 8, 14, 16 and 18 are selected as better experiments. The COPRAS procedure is incorporated to find the ranking of these experiments. The COPRAS grade calculations are done as discussed in section 2.3.3 and they are presented in table 11.
From table 11, it is noted that the experiment no. 16 produces optimal results and they are the outcome of 18 experiments only. But totally 243 (five parameters with three levels) combinations are possible. To find any other better results beyond the 18 experiments, the main effects of COPRAS grades are determined (table 12). The optimal combinations obtained from table 12 are observed as A3, B1, C2, D2 and E3. These combinations are not yet tested and they are also not available in L18 also.

Conclusions
A novel hybrid multi objective optimization technique has been presented in this article to optimize the process parameters in 3D printing. Taguchi method has been applied in this work for designing the experiments and the process parameters for maximizing the mechanical properties have been optimized using hybrid DT-FAHP-COPRAS methodologies. Among the several parameters considered, it is understood that the raster angle, raster width and the layer thickness are the most influencing parameters to obtain better mechanical properties. Finally, one more experiment called as confirmation experiment is conducted by utilizing the optimal parameters to ensure the results. It has been found that the process parameters under consideration have good effects on the strength of the built models. Hence it can be concluded that the proper selection of these process parameters will lead to the fabrication of stronger prototypes.

Data availability statement
No new data were created or analysed in this study.

Conflict of interest
The authors have no conflict of interest.

Funding statement
No funding assistance is availed related to this article.