Integrating Deep Learning into CAD/CAE System: Generative Design and Evaluation of 3D Conceptual Wheel

Soyoung Yoo; Sunghee Lee; Seongsin Kim; Kwang Hyeon Hwang; Jong Ho Park; Namwoo Kang

Source-linked AI summary

Integrating Deep Learning into CAD/CAE System: Generative Design and Evaluation of 3D Conceptual Wheel

Soyoung Yoo, Sunghee Lee, Seongsin Kim, Kwang Hyeon Hwang, Jong Ho Park, Namwoo Kang

arXiv:2006.02138v3May 25, 2020cs.GR

TL;DR

Conceptual CAD/CAE design needs rapid evaluation of many candidates, but generating CAD models and CAE data is costly. This paper presents a seven-stage deep-learning framework that generates 3D wheel CAD designs and predicts engineering performance from 2D designs. The road-wheel case study supports using AI to review many concepts and select candidates for detailed design, while retaining substantial upfront simulation and manufacturing-constraint limitations.

Problem
Conceptual design lacks an efficient way to create sufficient CAD models and CAE results for deep-learning surrogate evaluation across many candidates.
Method
The framework integrates 2D generative design, dimensionality reduction, latent-space DOE, CAD/CAE automation, transfer learning, and visualization for road-wheel concepts.
Results
The framework automatically generates numerous feasible 3D CAD designs and predicts CAE performance from 2D designs for conceptual wheel evaluation.
Takeaways & Limitations
Industrial designers and engineers can review generated CAD models with AI-estimated engineering results and select conceptual candidates for detailed design.
Takeaways & Limitations
The approach requires substantial upfront CAE data collection, including 1,006 simulations in this study, and does not address sufficient manufacturing constraints.

Abstract

from arXiv · show

Engineering design research integrating artificial intelligence (AI) into computer-aided design (CAD) and computer-aided engineering (CAE) is actively being conducted. This study proposes a deep learning-based CAD/CAE framework in the conceptual design phase that automatically generates 3D CAD designs and evaluates their engineering performance. The proposed framework comprises seven stages: (1) 2D generative design, (2) dimensionality reduction, (3) design of experiment in latent space, (4) CAD automation, (5) CAE automation, (6) transfer learning, and (7) visualization and analysis. The proposed framework is demonstrated through a road wheel design case study and indicates that AI can be practically incorporated into an end-use product design project. Engineers and industrial designers can jointly review a large number of generated 3D CAD models by using this framework along with the engineering performance results estimated by AI and find conceptual design candidates for the subsequent detailed design stage.

1. Introduction

The study addresses the difficulty of creating enough CAD models and CAE results for deep-learning evaluation during conceptual design. It proposes combining generative design, CAD/CAE automation, and deep learning to generate and evaluate 3D wheel concepts.

Conceptual design requires surrogate models that quickly evaluate many candidates, but engineers must first create CAD models and collect CAE results.
Generative design addresses data shortage by automatically exploring constrained design spaces, including through topology optimization and deep learning.
The study extends a prior 2D wheel generative-design method to 3D wheel design for an industrial application.
The proposed framework combines seven stages: 2D generation, dimensionality reduction, latent-space DOE, CAD automation, CAE automation, transfer learning, and visualization and analysis.
AI-generated CAD models and engineering-performance estimates allow industrial designers and engineers to select conceptual candidates for detailed design.

2. Related Work

Related work develops deep-learning representations for 3D CAD and engineering-performance prediction, while generative design helps create diverse training data. The paper positions its framework at the intersection of these needs.

Deep-learning models for 3D CAD commonly use voxel, point-cloud, or view-based representations, each requiring preprocessing of high-dimensional geometry.
Prior engineering-design studies use 3D CAD data and finite-element results to predict performance such as aircraft aerodynamics.
Collecting sufficient 3D CAD data in a target product domain remains a bottleneck for deep-learning-based CAD/CAE research.
Generative design automatically explores constrained designs and can provide initial concepts and inspiration during conceptual design.
Topology optimization explores material layouts and diverse designs through alternative initial designs, optimizers, and filtering methods.
Deep generative design combines topology optimization with deep learning to produce designs that balance engineering objectives with similarity to market products.
CNNs learn shape patterns, autoencoders reduce dimensionality, and transfer learning adapts pretrained feature extractors to smaller related datasets.

3. Deep CAD/CAE Framework

The deep CAD/CAE framework automates conceptual wheel generation and performance evaluation through seven connected stages. It uses 2D designs to create 3D CAD data and predict engineering results.

The framework has seven stages: 2D generative design, dimensionality reduction, latent-space DOE, 3D CAD automation, CAE automation, transfer learning, and visualization and analysis.
Stage 1. 2D generative design: Stage 1 uses topology optimization and deep learning to generate 16,689 2D disk-view wheel designs from commercial wheel references.
Stage 2. Dimensionality reduction: Stage 2 maps 128 × 128 wheel images into a 128-dimensional latent space with a convolutional autoencoder.
Stage 3. DOE in latent space: Stage 3 applies Latin hypercube sampling to draw 1,030 wheel designs from the latent space for CAD creation.
Stages 4–5. CAD and CAE automation: Stages 4 and 5 automate 3D CAD construction and modal-analysis data collection for lateral-mode natural frequency.
Stages 6–7. Transfer learning and analysis: Stage 6 predicts natural frequency and mass from 2D wheel designs using a CNN with transfer learning, while Stage 7 visualizes latent-space relationships and Grad-CAM explanations.

4. Design Generation (Stages 1 to 4)

The generation pipeline combines reference-guided topology optimization, filtering, and deep generative design to create varied 2D wheel topologies. Latent-space sampling then supports representative design exploration.

Stage 1 combines topology optimization and deep learning to generate many 2D disk-view wheel designs.
The optimization minimizes structural compliance while penalizing L1 distance from a reference wheel design.
The similarity weight λ controls the trade-off between matching the reference design and minimizing compliance.
Density filtering with Heaviside projection supplies manufacturing-related feature control during topology optimization.
658 reference wheel images produced 82,250 topology-optimization designs, which were filtered to retain designs with different topologies and shapes.

4.2. Dimensionality Reduction (Stage 2)

Stage 2 reduces augmented 2D wheel images to a 128-dimensional latent space using a convolutional autoencoder, then verifies that the representation preserves wheel features and supports smooth interpolation.

Data: Data augmentation increased the 2D wheel dataset from 16,678 to 166,812 images by randomly rotating wheels across 360°.The augmentation reflects rotational equivalence and improved deep-learning performance.
Architecture: A convolutional autoencoder maps each 128 × 128 wheel image to 128 latent dimensions and reconstructs the image through a decoder.The encoder and decoder use convolutional layers, with pooling in the encoder and upsampling in the decoder.
Training: The autoencoder minimizes pixel-wise reconstruction error using mean squared error between input and output images.The loss compares each input pixel x_i with its reconstructed value x̂_i.
Testing: Reconstructions of 96 unseen manufacturer wheel images were similar to their inputs, indicating that the latent space represented wheel-shaped features.The model also reproduced a central hole in a test image where no hole was present, reflecting a pattern learned from training data.
Testing: Interpolated wheel designs changed gradually between two different endpoint wheels while retaining features of both, indicating a continuous latent space.Ten equally spaced latent vectors were decoded between the endpoint representations.

4.3. DOE in latent space (Stage 3)

Stage 3 samples wheel designs in the autoencoder’s latent space, whose learned feature distribution provides a more representative design-exploration space than the original image dimensions.

Sampling: Latin hypercube sampling from a 128-dimensional multivariate normal latent distribution generated 3,000 sampled vectors for wheel-design exploration.The distribution used the training data’s latent-space mean and covariance, and LHSnorm produced good wheel-shaped images.
Comparison: Latent-space sampling achieved an average L1 distance of 0.5693, compared with 0.0225 for sampling in the original 128 × 128 image space.The larger latent-space distance indicates that its samples were more widely distributed for design exploration.
Comparison: The comparison found that DOE in the original image space could not represent the 2D wheel design space well.The study therefore used latent-space DOE to obtain more varied, representative designs.

4.4. 3D CAD Automation (Stage 4)

Stage 4 converts generated 2D wheel images and selected cross-sections into automatically generated 3D CAD models. Image processing extracts and organizes spoke geometry before Fusion360 automation creates the final wheel body.

Stage 4 overview: Stage 4 uses image processing, coordinate grouping, and Fusion360 automation to create 3D CAD models from 2D wheel designs.The process includes edge smoothing and sharpening, edge extraction, coordinate conversion, grouping, and automated modeling.
Image processing: Antialiasing smooths 128 × 128 pixel wheel images before edge detection identifies boundaries using horizontal and vertical brightness gradients.The Sobel operator is used for edge extraction, and rim and hub edges are removed so only spoke edges remain.
Data processing: Adjacent spoke-edge coordinates are sorted and grouped by distance, then filtered to produce smooth spline curves without tortuous or noisy geometry.Groups larger than 20 points are reduced, while groups with 3 or fewer points are treated as noise and deleted.
3D modeling: 3D modeling combines disk-view spoke shapes with selected spoke and rim cross-sections to generate wheels sharing a representative 18-inch, 7.5j rim configuration.The cross-sectional geometry is extracted from a selected CAD model and reused while spoke shapes vary.
3D modeling: The automated workflow sketches, revolves, extrudes, cuts, and joins bodies before saving each completed wheel as a universal .stp CAD file.Fusion360’s Python API performs the modeling sequence, including spoke creation, lug-hole cuts, rim generation, and final body joining.
Output: The resulting workflow produces automatically generated 3D wheel CAD models for subsequent engineering evaluation.Examples of the generated models are shown in Figure 16.

5. Design Evaluation (Stages 5 to 7)

Stages 5 to 7 automate modal-analysis evaluation, deep-learning prediction, and visualization of generated wheel designs. The framework predicts engineering performance from 2D designs and helps identify high-stiffness candidates while exposing shape–performance relationships.

CAE automation: Free-free modal analysis obtains natural frequencies and mode shapes, with lateral mode 11 selected to evaluate spoke-shape performance.Stiffness is calculated from natural frequency and mass, while manufacturers use lower stiffness bounds as design constraints.
CAE automation: CAE automation imports 3D CAD models, generates second-order tetrahedral meshes, and runs modal analysis using reference aluminum-wheel material properties.The automated mesh size is constrained to no larger than 6 mm.
Deep-learning evaluation: Transfer learning and ensemble modeling predict modal-analysis results and mass from 2D designs despite limited training data.The approach combines a pretrained convolutional autoencoder and DNN, while data augmentation expands 1006 designs into 10,060 training samples.
Deep-learning evaluation: TL_CAE_Ensemble achieves the highest predictive performance in RMSE and MAPE among the four trained models.TL_CAE improves RMSE and MAPE over TL_VGG16, and the ensemble averages nine frequency predictions and five mass predictions.
Testing: 0.66 s is the average time required to evaluate a wheel concept using predicted frequency and mass on an NVIDIA TITAN Xp 4.8 GB GPU.The predicted results support stiffness-based ranking and elimination of unsatisfactory designs before detailed design.
Feature visualization: Latent-space visualization groups wheel shapes and natural-frequency values, showing that similar frequencies cluster and thicker spokes correspond to higher frequency.The framework embeds wheel data into latent space, reduces it with T-SNE, and uses K-means groups for visualization.
Feature visualization: Visualization enables intuitive selection of high-performance design clusters and visual checking of CAD/CAE prediction reliability.It also supports analysis of shape features associated with better engineering performance.
Grad-CAM: Grad-CAM identifies the wheel center as an important region affecting frequency, linking higher frequency to a more filled center in lateral mode.The interpretation is consistent with the largest displacement occurring at the wheel center in the lateral mode shape.

6. Discussion and Conclusion

The framework combines generative design, CAD/CAE automation, and deep learning to generate and evaluate conceptual 3D wheel designs. Its practical scope is demonstrated alongside trade-offs involving design diversity, manufacturing constraints, computational cost, and data requirements.

Generation and evaluation: The generation phase balances compliance minimization with similarity to a reference design to produce stiff, wheel-like 2D designs.This formulation supports subsequent 3D CAD generation but can constrain diversity when force conditions are not varied.
Generation and evaluation: The evaluation phase can filter generated wheels using multiple engineering analyses, although this study demonstrates modal analysis as one example.Potential evaluations include rim and disk stiffness, impact strength, and aerodynamics.
Generation and evaluation: Generating designs before evaluation allows product teams to select candidates according to differing engineering-performance priorities and product concepts.The authors frame this ordering as easier to use industrially because evaluation standards and performance priorities vary by product concept.
Limitations and future work: The current 2D-mediated approach is retained because direct 3D generative design faces difficult reference-data, convergence, dimensionality, and computational-cost problems.Future work targets 3D-input learning and generative design without using 2D images.
Comparison with previous design approach: The framework uses latent-space DOE and 2D-to-3D automation to create more varied, market-like designs and more 3D models than conventional 3D topology optimization.Conventional topology optimization instead tends toward unfamiliar shapes, while the proposed approach integrates seven stages from 2D generation through visualization and analysis.
Limitations: The study requires 1,006 CAE simulations for training, does not sufficiently address manufacturing constraints, and uses reference-based filtering to achieve symmetry.The authors note that this front loading may improve efficiency across subsequent product design projects, but it remains a limitation for rapid evaluation.
Conclusion: The framework automatically generates numerous feasible 3D CAD models and evaluates engineering performance, supporting joint designer–engineer review of conceptual candidates.Its deep learning surrogate model also relates geometry to engineering performance for conceptual wheel designs.

Declaration of interests

The authors report no known competing financial interests or personal relationships that could have influenced the work. The code is not disclosed because Hyundai Motor Company owns it.

The authors declare no known competing financial interests or personal relationships that could have influenced the reported work.
The paper’s code cannot be disclosed because it is the property of Hyundai Motor Company, which funded the study.

Appendix A: Data augmentation effect on autoencoder

The appendix reports that data augmentation improves autoencoder learning and reconstruction quality compared with training without augmentation.

The loss value is relatively small when data augmentation is used compared with the case without data augmentation.
Reconstructions without data augmentation have relatively poor image quality compared with the images shown in Figure 6.

Appendix B: Detailed algorithm for sorting and grouping points

The appendix describes a nearest-point grouping algorithm that uses a distance threshold, then reduces sorted points into closed spline curves recognized as CAD surfaces.

Point sorting and grouping: The algorithm initializes a point, repeatedly selects and removes the nearest point, and stores points into groups.
Point sorting and grouping: Points join the current group when their distance from the fixed point does not exceed the threshold; otherwise, a new group is created.
Threshold selection: The final distance threshold was chosen as five after tests found complete spoke-group separation at thresholds three and above.
Spline construction: Sorted group points are subsampled and closed by repeating the first coordinate, allowing CAD software to recognize the curve as a surface.