KNOWLEDGE-BASED ADAPTATION OF PRODUCT AND PROCESS DESIGN IN BLISK MANUFACTURING

Early and efficient harmonization between product design and manufacturing represents one of the most challenging tasks in engineering. Concepts such as simultaneous engineering aim for a product creation process, which addresses both, functional requirements as well as requirements from production. However, existing concepts mostly focus on organizational tasks and heavily rely on the human factor for the exchange of complex information across different domains, organizations or systems. Nowadays product and process design make use of advanced software tools such as computer-aided design, manufacturing and engineering systems (CAD/CAM/CAE). Modern systems already provide a seamless integration of both worlds in a single digital environment to ensure a continuous workflow. Yet, for the holistic harmonization between product and process design, the following aspects are missing:


PRODUCT AND PROCESS DESIGN IN TURBOMACHINERY MANUFACTURING
For decades, traditional product development methods considered product design (development) and process design (manufacturing) as sequential steps (see FIGURE 1). At a certain handover point in product development, the completed product design was transferred to process design in order to plan and execute the manufacturing process. However, insufficient consideration of manufacturability during product design often resulted in non-manufacturability of the product, cost intensive design adaptations in a late phase of product development or uneconomic manufacturing processes.

FIGURE 1: CONVENTIONAL PRODUCT DEVELOPMENT
In order to overcome the described deficit of a sequential product and process design, both research and industry have investigated different product development methods. Such methods are for example the "simultaneous engineering" approach (also known as concurrent engineering), which aims for a parallelization of several steps along product development [1]. Especially in turbomachinery, manufacturing complex products and extremely long-term product development processes require adapted approaches. An exemplary approach is the "integrated development, manufacturing and maintenance" (IP3E) approach of MTU Aero Engines AG, which as well integrates the highly relevant series production and maintenance phase of the industry (see FIGURE 2) [2].

FIGURE 2: CONCEPT OF SIMULTANEOUS ENGINEERING
Modern product development methods mostly focus on organizational tasks and heavily rely on the human factor for the exchange of complex information across different domains, organizations or systems. As a result, product development in turbomachinery manufacturing is still characterized by extensive iterative testing procedures between development and manufacturing. Optimal product designs, which provide a maximum of functionality while ensuring a high manufacturability are not determined systematically. These circumstances become even more important considering the increasing complexity of future jet engine designs.
To achieve a fully harmonized product and process design a systematic coupling of the advanced software tools of both phases is required. FIGURE 3 shows the most relevant software tools in product and process design for complex turbomachinery components. Especially in the field of process planning and analysis, numerous studies have been conducted, which individually address process verification and simulation as well as quality prognosis [3][4][5][6][7][8][9][10][11].

FIGURE 3: SOFTWARE IN PRODUCT AND PROCESS DESIGN
Modern product life cycle (PLM) systems already provide a seamless integration of CAD/CAM/CAE in a single digital environment to ensure a continuous workflow. Yet, for the holistic harmonization between product and process design, the following aspects are missing:  The digital environment does not provide a complete and data consistent digital twin of the component; this V007T17A012-2 Copyright © 2021 by ASME; reuse license CC-BY 4.0 applies especially to the process design and analysis environment  Due to the lack of process and part condition data in the manufacturing environment an adaptation of product and process design for a balanced functionality and manufacturability is hindered  Systematic long-term data analytics across different product and process designs with the ultimate goal to transfer knowledge from one product to the next and to accelerate the entire product development process is not considered

PROBLEM COMPLEXITY VS. DATA AVAILABILITY IN BLISK MANUFACTURING
Modern turbo engines make extensive use of so-called blade-integrated disks (blisk). Blisks are integral rotor components, which combine disk and blades within a single component. The production of such components represents one of the most challenging tasks in turbomachinery manufacturing. Complex, thin-walled and high aspect ratio blade geometries made from hard-to-cut materials (Ti-/Ni-alloys), as well as extremely tight tolerances put highest demands on product and process design.
Due to the geometrical complexity of blisks and the applied multi-axis machining processes, the CAM-system is the central element of process design. Concerning the high demand for blisks in the aerospace sector, CAM-system providers have developed customized machining strategies, so-called "multiblade strategies". These appear to be comparatively highly automated and supposedly less complex. The number of operation parameters of conventional machining strategies is about 100 to 250, those of multi-blade strategies about 50 (counting in the CAM-system). Usually it is not necessary to adjust all parameters for a successful process planning. Experience shows that the number of parameters that are actually relevant to the process or part quality is about 15 (counting in the CAM-system). If the user intends to test only three values per parameter, there are about 14.3 million (3 15 = 14,348,907) possible solutions (tool and cutting parameters not included). During process design for a new component, the CAM-planner is only able to search a fraction of this multi-dimensional solution space for optimal parameter settings in time consuming and iterative manual optimization procedures. During this process, he is bound to the low computing power, insufficient process analysis capabilities and missing knowledge management of today's CAM-systems.
The transfer of knowledge from one product to the next can potentially enhance repetitive optimization tasks. Such transfer of knowledge from related tasks is particularly important for scenarios where the amount of available training data (simulation results, labelled measurements) is small compared to the intrinsic complexity of the problems at hand (such as blisk manufacturing). For such scenarios, it is of high practical relevance to study algorithmic approaches that directly address the "small data -high complexity" issue, such as multi-task learning [12], transfer learning [13,14], or evolutionary algorithms (EA) and artificial intelligence (AI) that incorporate elements of facilitated variation [15].
The consideration of engineering problems by means of evolutionary algorithms (EA) or artificial intelligence (AI) has been the subject of scientific studies for quite some time. Evolutionary algorithms are strongly oriented on the evolutionary mechanisms of nature (selection, recombination, mutation, evaluation) and are supposed to solve optimization problems in large solution spaces. Artificial intelligence, on the other hand, uses functional mechanisms of the human brain and investigates complex relationships or tasks that cannot be fully understood by humans. Today, machine learning in particular is a key technology of artificial intelligence and is making promising progress in the field of image data recognition, for example in medical technology or finance [16]. However, an industrial use of EA or AI within commercial CAM systems is not state of the art today [17]. However, some researchers are already showing the benefits of these methods in the context of manufacturing technology [18,19].
The first deficit is the complexity and relevance of the addressed optimization problems. Considering current research, it becomes obvious that the optimized target variables are usually of purely geometric nature (e.g. minimization of non-productive machine movements or tool change movements etc.). Only few works focus on the optimization of technological process and part condition variables (e.g. process forces, residual stresses, surface quality etc.). Here again the circumstance of a missing holistic technological process analysis within the CAM-system in the absence of integrated coupled simulation models is the root cause. Furthermore, the limit of the average available computing power plays a role, which can make the scientific investigation of specific optimization issues difficult or even impossible.
A second deficit is the singular use of the different methods of EA and AI (evolutionary algorithms, artificial neural networks, swarm intelligence etc.). Only in a few cases a combination of methods (e.g. combination of evolutionary algorithms with artificial neural networks) is used to achieve a faster and more efficient optimization process. However, this area in particular offers a high potential for research and development, for example by replacing random undirected optimization with fast targeted optimization.
A third deficit is the almost exclusive consideration of simple prismatic workpieces and comparatively simple 2-or 3axis machining. This reduces the amount of input variables (especially the process parameters) to a minimum. Simultaneous multi-axis machining processes for complex workpieces in particular place significantly higher demands on the optimization mechanisms to be developed due to the higher multidimensionality and the strong interactions of individual parameters.

CONCEPT FOR THE EXPLORATION AND KNOWLEDGE-BASED ADAPTATION OF PRODUCT AND PROCESS DESIGN
In order to overcome the described small data -high complexity issue and to establish an environment for the

V007T17A012-3
Copyright © 2021 by ASME; reuse license CC-BY 4.0 development of improved optimization algorithms an exploration concept has been set up. The exploration concept has been implemented as a standalone software executable (test bench), which couples product design (CAD) and process design (CAM), along with individual process analysis (CAE), evaluation and adaptation functionalities for the process design environment (see FIGURE 4).

FIGURE 4: EXPLORATION AND ADAPTATION CONCEPT
The exploration concept is initially focused on the exploration of large solution spaces based on random parametric variations of product and process design and the evaluation of different product designs with regard to manufacturability through integrated process analyses. The process design (CAD) comprises a parametrized blisk model implemented in a commercial CAD-core, derived from a compressor blisk design with challenging manufacturing features [20] ( see FIGURE 5).

FIGURE 5: PARAMETRIZED BLISK MODEL
The model allows the parameter-based variation of the number of blades, blade height, blade twist, blade thickness, tip forward sweep and other geometrical aspects. Based on the parametrized blisk model and a random parameter variation, the product design is able to create a large variety of blisk designs, each represented by a set of drive surfaces (shroud surface, blade surface, collision surfaces and hub surfaces).
The process design (CAM) receives the drive surfaces (tessellated boundary representations BREP), along with a cutting tool definition (fixed) and operation parameters for the toolpath calculation. An integrated commercial 5-axis CAMcore subsequently performs the required toolpath calculations. A random variation of the operation parameters allows the calculation of a large variety of different operation variants on a single blisk geometry. Exemplary parameters to be adjusted are the beginning and end value of a machining operation, lead and tilt angle, interpolation tolerance, side step and other operation parameters. All toolpaths from the operation list are calculated, representing the entire machining process for the blade area (see FIGURE 6).

FIGURE 6: PARAMETRIZED TOOLPATH CALCULATION
The process analysis (CAE) module receives the calculated toolpath, cutting tool definition, as well as raw and finish part geometry as input. It calculates macroscopic engagement data based on a numeric multi-dexel model [21]. Subsequently an analytical model calculates the microscopic engagement data (uncut chip geometry) from the macroscopic engagement data [22]. The uncut chip geometry serves as input for a dualmechanistic cutting force simulation in combination with empirically determined calibration factors [23]. Based on the engagement and force data a number of analytical, (semi-) empirical and numerical models calculate further physical properties of the process and the workpiece such as tool and workpiece deflections etc. (see FIGURE 7). Conclusively, the

V007T17A012-4
Copyright © 2021 by ASME; reuse license CC-BY 4.0 system saves all process and part condition data, representing the digital twin of the process design environment.

FIGURE 7: PROCESS ANALYSIS AND DIGITAL TWIN
Through extensive parallelized simulation runs, the system is able to create large amounts of data with regard to complex optimization problems under varying blisk designs. Exemplary optimization issues comprise the toolpath (even NC-point distribution, smooth vector orientations etc.), the process (low and homogenous cutting forces, low tool wear etc.) and the part itself (geometrical deviations, block transitions, surface roughness etc.). Data analysis, either manually or by the help of machine learning, is performed in order to identify correlations and causalities between the setting parameters and the resulting process and part quality.
The resulting insights allow for the development of smarter optimization algorithms, which enable a faster and better process design in industrial application. Key aspect of such smarter optimization algorithms is the transfer of knowledge from one product design to the next in order to limit the required amount of iterations during process design (elements of facilitated variations). The reduced amount of required iterations allows the integration even of complex simulation models in the later CAM-system for industrial application.
Within the presented paper the investigations are limited exclusively to the process design. FIGURE 8 shows the vision of integrating elements of facilitated variation both in product and process design. Both areas might ultimately operate on the same digital twin, allowing e.g. the CFD-simulation of blade geometries, which carry geometrical deviations from manufacturing. Simultaneous fitness evaluation of the product design in terms of product use and manufacturing would ultimately allow for the systematic determination of optimal product designs, which provide a maximum of functionality while ensuring a high manufacturability.

PROCESS DESIGN TASK I -NUMBER OF BLOCKS
Modern blisk designs are characterized by high aspect ratio blade geometries. Machining of such geometries is usually executed block-wise from top (shroud) to bottom (hub). Block-wise machining limits the free overhang of the blade in the machined area. Thereby it reduces the static-dynamic workpiece and tool deflection, as well as the tendency for process instabilities or chatter and enhances the geometrical accuracy and surface quality of the blades (see FIGURE 9).

FIGURE 9: BLOCK-WISE MACHINING OF BLISK GEOMETRY
Initially during process design, the CAM-planner needs to determine a suitable number of blocks for the individual blisk design. This working task is performed entirely based on experience. The CAM-system provides no suitable process analysis functionalities to support the working task. The extensive application, e.g. of a static-dynamic deflection simulation, in the industrial CAM-system is hindered due to the high required computation time of such simulations. Due to the described circumstances the case represents a "high complexity -small data" issue. It requires the setup of a suitable exploration system, the analysis of the problem based on extensive simulation runs and the subsequent derivation of automated and knowledge-based optimization approaches for industrial application (facilitated variation). The following results serve as an initial exploration of the optimization case.

V007T17A012-5
Copyright © 2021 by ASME; reuse license CC-BY 4.0 The objective was to determine a suitable number of blocks for three different blisk designs based on a static deflection simulation. The initial compressor blisk design from chapter 3 serves as geometry A with a blade height hB of 50 mm. Geometry B has a blade height of 60 mm (+10 mm), while geometry C has a blade height of 40 mm (-10 mm). In order to limit the required computation time of the exploration, static-dynamic simulations based on an integrated FEM-simulation were neglected. The required radial force Fr was calculated based on the cutting force simulation described in chapter 3. The applied calibration factors were determined for the specific cutting tool and workpiece material Inconel 718 (AMS 5663), shown in TABLE 1. The finishing tool was approximated as an euler bernoulli beam with a circular cross-section and a length of Lt = 52 mm, a diameter of dt = 6 mm and an elastic modulus of Et = 600 GPa (solid carbide). The blade was approximated as an euler bernoulli beam with a rectangular cross-section of width bb = 22.5 mm, a thickness of thb = 1.5 mm and an elastic modulus of Eb = 204 GPa (Inconel 718). The varying cantilever length Lw of the blade was calculated as the difference between the current contact point information of the finishing toolpath and the lowest contact point of the semi-finishing operation in each block. For the calculation of the second moment of area, as well as the tool and workpiece deflection the following equations were used: Simulations of the radial force Fr, the tool deflection ẟt, the workpiece deflection ẟb and the total deflection ẟtotal (sum of tool and workpiece deflection) were carried out on the three different blisk designs A-B-C. The number of blocks was varied from 1-7 blocks. FIGURE 11 shows an example of the results of the simulations for a two and four block strategy. The simulation results are plotted spatially along the tool path of the finishing process. The number of blocks has no influence on the simulated radial force Fr and the tool deflection ẟt. In both cases, the maximum radial force Fr is 25 N and the maximum tool deflection ẟt is 30 µm. The number of blocks has a strong influence on the workpiece deflection ẟw of the blade. In case of a two block strategy the maximum workpiece deflection ẟw in both blocks is >> 20 µm. In contrast, a maximum workpiece deflection ẟw of > 20 µm appears in the uppermost area of all blocks for the four block strategy. In qualitative terms, this observation also applies to the total deflection ẟtotal.
The simulated tool, workpiece and total deflections in relation to the three different blisk geometries (A-B-C) and numbers of blocks (1-7) are shown in FIGURE 11. The total deflection with regards to the number of blocks can be described by a regressive course. A limit value of 0.05 mm was defined as a tolerable total deflection ẟtotal. ẟtotal falls below the limit in case of a three block strategy for geometry C, a four block strategy for geometry A and a five block strategy for geometry B.

V007T17A012-6
Copyright © 2021 by ASME; reuse license CC-BY 4.0 process parameter number of blocks. In industrial application, testing of different numbers of blocks is extremely time consuming and virtually not feasible (mainly due to the required high-precision engagement and force simulation). Execution of the presented simulations based on the presented exploration system running on a high-performance workstation took several days. Accordingly, the exploration data will be used in the future to derive knowledge-based optimization approaches which reduce the required iteration steps for the same optimization procedure (facilitated variation) and ultimately allow the industrial integration within the CAM-system.

PROCESS DESIGN TASK II -BLOCK TRANSITION
Block-wise machining enhances the overall geometrical accuracy and surface quality of blades by reducing staticdynamic deflections and the tendency for process instabilities or chatter. However, it can result in discontinuities at the transitions between different blocks. A major root cause for discontinuous block transitions is the choice of inadequate block end values in finish machining. Before finish machining, the purpose of semifinish machining is to create a homogenous material offset situation in the relevant blade area. When choosing an inadequate (too high) block end value for finish machining the cutting tool might engage with the remaining inhomogeneous rest-material below the semi-finished area, or even the raw material in the following block (see FIGURE 12). That results in a significant rise of the process forces in that area, causing high tool deflections and leaving behind discontinuous block transitions. In order to reduce or avoid discontinuous block transitions, the CAM-planner needs to determine suitable shift values (percentage of blade height as a decimal number) between semifinishing and finishing for all blocks within the machining operation of the individual blisk design. In order to limit the free overhang of the blade at the beginning of the next finishing operation these shift values need to be as small as possible. The CAM-system provides no suitable process analysis functionalities to support the working task (for example an engagement or force simulation). In addition and analog to the previous case the extensive application of an engagement simulation for process analysis in the industrial CAM-system is hindered because of the required computation time. Due to the described circumstances the case as well represents a "high complexity -small data" issue. Again, the following results serve as an initial exploration of the process design task. With a shift value of 0.012 a small peak occurs at the end of each force signal, respective the toolpath. If the shift value is further reduced to 0.008, this peak rises sharply for all geometries due to engagement of the cutting tool with the inhomogeneous restmaterial below the semi-finished area. In this case 0.016 represents the optimal shift value for all three geometries within block 2. FIGURE 14 shows the simulated resulting forces that occur for different shift values in block four. It is noticeable that the geometries show different behaviors. For geometry A, a force peak only occurs for the smallest shift value 0.008. For geometry B, the force peak already occurs for a shift of 0.012, while for geometry C a force peak occurs already for a shift value of 0.016. Accordingly, optimal shift values represent 0.012 for blisk   To determine the optimal shift value for each block in each geometry, we define the force increase factor as the ratio of the maximum force at the end of the feed travel ( ) to a baseline value of the force in the middle of the feed travel ( ): In case of a too small shift value, a peak of the force is observed at the end of the feed travel, as described in the previous paragraphs, resulting in > and > 1. We define as the maximum force value in the last 10 % of the feed travel and as the maximum force value in the tool path region from 40 % to 60 % of the feed travel.
The optimal shift value can reliably be determined by calculating over a range of shift values. FIGURE 15 shows R across the different shift values of the simulation and the five different blocks. As can be seen, R decreases with increasing shift until it approximates 1 at the optimal shift value. For the blocks one and two, the optimal shift values for all geometries are 0.016. In case of block three, four and five, the optimal shift value highly depends on the geometry. For example, geometry C has the highest R values overall and compared to geometries A and B an optimum value is only achieved at higher shift values. As for process design task one, the simulations of process design task two show that R is highly dependent on the blisk design, as well as the process parameter shift value. Again, the testing of different shift values is extremely time consuming and virtually not feasible in industrial application. Exploration data will be used in the future to derive knowledge-based optimization approaches which reduce the required iteration steps for the optimization procedure (facilitated variation) and ultimately allow the industrial application within CAM.

CONCLUSION
The presented exploration system is able to create large amounts of data for complex optimization issues in blisk manufacturing across different product and process designs. This exploration enables the identification of critical optimization issues and to overcome the "small data -high complexity issue", described in chapter 2 of the paper. The presented process design tasks show a high dependency of the process design from the product design, therefore representing critical starting points for the development of "facilitated variation" algorithms. The next steps are the following:  Extension of the exploration system by means of further process simulation capabilities in order to enhance the completeness of the digital twin  Further exploration and validation in terms of additional simulation runs and machining trials for the presented process design tasks I/II  Addition and exploration of further process design tasks  Copyright © 2021 by ASME; reuse license CC-BY 4.0  Development of facilitated variation algorithms and proof-of-concept of a significant reduction of simulation efforts for optimization  Recommendations for a transfer of the presented approach to other product types or industries