Generative AI-Driven Parametric Design Optimization Framework for Vehicle Closure Systems
Authors: Dongeon Lee, Yurim Lee, Jiyong Kim, Sanghyeon Kim, Sanghyun Lee, Seungyeon Shin, Jeahong Kim, Namwoo Kang
Organizations: Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology, Daejeon, 34051, Republic of Korea · Division of Future Vehicle, Korea Advanced Institute of Science and Technology, Daejeon, 34141, Republic of Korea · Hyundai Motor Company, Hwaseong, 18280, Republic of Korea · Department of Mechanical Engineering, Korea Advanced Institute of Science and Technology, Daejeon, 34141, Republic of Korea · Narnia Labs, Daejeon, 34051, Republic of Korea
This study presents a generative AI-driven parametric design optimization framework integrating performance-oriented generation with editable computer-aided design (CAD) and downstream computer-aided engineering (CAE) analysis for industrial vehicle closure systems. The framework is demonstrated on an automotive tailgate assembly with complex geometry, coupled components, and competing performance and feasibility requirements. A production-oriented CATIA model is restructured into a reference-based parametric template with a 13-dimensional design representation. An automated CAD/CAE workflow generates parametric models, performs structural analysis, and constructs a database of mass, maximum displacement, and maximum von Mises stress. Gaussian process regression surrogates predict performance, while diffusion and flow matching models generate design candidates. The design problem minimizes mass and maximum displacement, with stress assessed as an engineering feasibility constraint. Performance-guided generation using surrogate guidance and performance-aware training using a performance-aware loss incorporate engineering objectives, with sensitivity to guidance scale and loss weight evaluated. Generative search is compared with NSGA-II under matched evaluation budgets using a common surrogate model. Selected Pareto-region candidates are reconstructed in CATIA and re-evaluated through native CAE; all verified candidates are successfully regenerated and analyzed. Relative to the baseline, the knee-point design reduces mass by 7.1% and maximum displacement by 29.4%, while extreme verified solutions achieve maximum reductions of 10.3% and 37.3%, respectively. The results demonstrate performance-oriented multi-objective design exploration compatible with editable industrial CAD and native engineering verification.
Figures & tables
Figure 1: Overview of the proposed generative AI-driven parametric design optimization framework.
Figure 2: Conceptual comparison between independent absolute dimensioning and reference-based dimensioning. In the reference-based formulation, editable dimensions are expressed relative to reference locations so that dependent geometry follows changes in upstream features.
Figure 3: Examples of CAD restructuring for the tailgate case study, including reference-based dimensioning to maintain geometric dependencies during parameter updates and extended base-surface construction to provide sufficient geometric margin for robust design variation.
Figure 4: Automated CAD/CAE workflow for constructing the design–performance database, including parametric CAD regeneration, CAE analysis, response extraction, and separate logging of CAD and CAE failures.
Figure 5: Generation processes of the two generative model families on the normalized design-variable space: (a) reverse-time denoising in the diffusion model and (b) forward-time integration of the learned velocity field in the flow matching model. Both models transform samples from a standard normal distribution into normalized 13-dimensional design-variable vectors.
Figure 6: Exploded view of the major structural and functional components of the automotive tailgate assembly used in the case study.
Design variable
Associated component
Geometric parameter
x1
Outer panel
Spoiler-region height
x2
Outer panel
Lower-section height
x3
Outer panel
Longitudinal length
x4
Outer panel
Spoiler-region width
x5
Back glass
Width
x6
Outer panel
Center-section width
Table 1: Definition of the 13 geometric design variables used in the tailgate case study.
Figure 7: Geometric locations of the 13 design variables selected for the tailgate case study. Absolute dimensional values and design-variable bounds are omitted for proprietary reasons.
Item
Adopted configuration
GPR surrogate models (Stage 4)
Input
13 normalized design variables
Response
Mass
Maximum displacement
Training / test split
151 training / 38 test designs
Covariance kernel
Matérn ( ν=5/2 ) + rational quadratic (ARD)
Matérn ( ν=1/2 ) (ARD)
Model selection
Repeated five-fold cross-validation
Table 2: Principal model architectures and training settings used in the case study.
Workflow step
Cases processed
Successful cases
Failed cases
Success rate
CAD generation
200
195
5
97.5%
CAE analysis
195
189
6
96.9%
Overall CAD/CAE workflow
200
189
11
94.5%
Table 3: Summary of automated CAD/CAE data generation outcomes for the 200 LHS designs.
Figure 8: Distributions of the normalized engineering responses obtained from the 189 retained CAD/CAE designs: (a) mass, (b) maximum displacement, and (c) maximum von Mises stress. The solid and dashed vertical lines indicate the mean and median, respectively.
Response
Test R2
5-fold CV R2
MAE
RMSE
MAPE
95% PI coverage
Mass
0.9998
0.9996 ± 0.0001
0.0051
0.0069
0.03%
97.4%
Displacement
0.8889
0.8505 ± 0.0452
0.2564
0.3443
5.81%
100.0%
Table 4: Predictive performance of the GPR surrogate models for mass and maximum displacement.
Figure 9: Hold-out parity plots of the GPR surrogate models for (a) mass and (b) maximum displacement using 38 test designs. The axes are normalized using the min–max ranges of the training data.
Figure 10: Training and validation losses of the adopted generative-model configurations at λ=0.05 : (a) diffusion and (b) flow matching. Solid and dashed curves denote the mean training and validation losses over five seeds, respectively; faint curves show the unsmoothed training losses, shaded bands indicate one standard deviation of the validation loss across seeds, and dotted vertical lines mark the mean epoch selected based on the minimum validation loss.
Family
Method
HV ↑
ND count ↑
Maximum spread ↑
Diffusion
Baseline
1.188 ± 0.110
6.4 ± 1.8
0.424 ± 0.220
Performance-guided generation only
1.384 ± 0.035
12.8 ± 1.6
0.591 ± 0.114
Performance-aware training only
1.231 ± 0.088
8.4 ± 1.7
0.594 ± 0.123
Combined
1.401 ± 0.031
13.4 ± 2.2
0.553 ± 0.094
Flow matching
Baseline
1.159 ± 0.046
6.6 ± 2.1
0.541 ± 0.175
Performance-guided generation only
1.233 ± 0.039
8.0 ± 2.6
0.552 ± 0.088
Table 5: Ablation results for performance incorporation at a surrogate-evaluation budget of N=500 .
Figure 11: Effect of performance incorporation on the Pareto distributions generated by the diffusion model at N=500 : (a) baseline, (b) performance-guided generation only, (c) performance-aware training only, and (d) combined. The generated designs and their nondominated fronts are shown together with the training data, with the NSGA-II front under the same evaluation budget included for reference.
Figure 12: Quantitative ablation results for diffusion and flow matching at N=500 : (a) hypervolume, (b) nondominated count, and (c) maximum spread.
Figure 13: Sensitivity of hypervolume to the performance-incorporation parameters: (a) guidance scale κ with λ=0.05 and (b) performance-aware loss weight λ with κ=0.5 .
Method
Evaluation budget, N
50
100
250
500
1000
2000
Hypervolume (normalized)
Diffusion – Combined
1.332 ± 0.035
1.345 ± 0.037
1.381 ± 0.028
1.401 ± 0.031
1.419 ± 0.023
1.429 ± 0.018
Flow matching – Combined
1.148 ± 0.061
1.195 ± 0.025
1.230 ± 0.024
1.260 ± 0.037
1.298 ± 0.023
1.316 ± 0.039
NSGA-II
1.117 ± 0.089
1.157 ± 0.076
1.294 ± 0.055
1.378 ± 0.043
1.439 ± 0.026
1.464 ± 0.021
LHS random search
1.047 ± 0.039
1.105 ± 0.027
1.158 ± 0.048
1.243 ± 0.047
1.260 ± 0.034
1.282 ± 0.020
Table 6: Hypervolume and nondominated count under matched evaluation budgets.
Figure 14: Budget-dependent comparison of the generative models, NSGA-II, and LHS random search. (a)–(c) Nondominated fronts of combined diffusion and NSGA-II across five seeds at N=100,250,500 , with the training data shown for reference. (d) hypervolume and (e) nondominated count as functions of the evaluation budget. The dashed line in (d) marks the approximate crossover between combined diffusion and NSGA-II.
Design
Normalized mass
Normalized displacement
Change from baseline
Structural feasibility ( σ≤σallow )
Pred.
CAE
Error (%)
Pred.
CAE
Error (%)
Δ Mass
Δ Disp.
Δ Stress
Baseline design
–
0.396
–
–
0.231
–
–
–
–
–
Minimum-mass
−0.327
−0.328
0.01%
0.098
0.001
20.5%
−10.3 %
−32.8 %
−20.4 %
Pass
Knee-point (balanced)
−0.105
−0.103
0.03%
0.025
0.025
0.1%
−7.1 %
−29.4 %
−17.9 %
Pass
Minimum-displacement
0.281
0.286
0.07%
−0.015
−0.030
3.4%
−1.6 %
−37.3 %
−31.1 %
Pass
Table 7: Native CAD/CAE verification of the three representative designs selected from the generated Pareto front, with the baseline design listed for reference. Objective values are normalized by the min–max range of the training split. Prediction errors are computed from the original physical units. Percentage changes are calculated from the direct CAE responses relative to the baseline design.
Figure 15: Native CAD/CAE verification of generated Pareto-region candidates in the normalized objective space. Training data and the baseline design are shown for reference, while circled generated candidates exceed the allowable-stress criterion. Three representative solutions are highlighted: minimum-mass, knee-point (balanced), and minimum-displacement.
Agentic AI has automated parts of scientific discovery, including paper generation, expert-level coding, therapeutic proposal, and autonomous experimentation. Complex physical engineering design remains a gap, because candidates must satisfy simultaneous constraints in fluid dynamics, solid mechanics, and structural stability. We introduce The AI Engineer, an agentic framework that couples large language models (LLMs) to deterministic engineering backends in a closed loop: natural-language requirements are converted into design-domain geometry and mesh; topology is optimized with bi-directional evolutionary structural optimization (BESO) coupled to the CalculiX solver; and member sizes are refined with particle swarm optimization (PSO) coupled to Zwind under offshore aero-hydro-servo-elastic load cases. To explore many designs without per-candidate certification cost, an Automated Reviewer scores each candidate on five dimensions (capacity, steel intensity, unit cost, constructability, and fatigue life) using piecewise-linear functions calibrated on 11 real floating-wind projects. Search terminates only when a candidate reaches a composite score S≥85 (grade A) with no subscore below 60. We validated this gate by submitting the top-scoring design to the China Classification Society (CCS) for Approval in Principle (AIP), which it passed; AIP is thus an external check that the reviewer tracks professional judgment, not the daily objective. The certified design outperforms the human-optimized TuQiang baseline, reducing steel mass and unit capital cost by 8.1% each while meeting all AIP criteria. This verification-closed regime, in which every proposal is judged by deterministic physics and codified limit states, distinguishes The AI Engineer from open-ended generative systems. Remaining limits include detailed design and fabrication-hard constraints.
Tianyi Yu, Chengxing Tao, Haoxuan Shen +8
State Key Laboratory of Ocean Sensing & Ocean College, Zhejiang University, Zhoushan, 316021, China · Zhejiang Province Key Laboratory of Offshore Civil Engineering and Materials & Ocean Academy, Zhejiang University, Zhoushan, 316021, PR China · Institute of Fundamental and Transdisciplinary Research, Zhejiang University, 310058, PR China +3
Recent advances in large language models (LLMs) have demonstrated significant potential in supporting engineering design tasks, including computer-aided design (CAD) automation. However, most existing LLM-based 3D CAD generation approaches primarily focus on geometric precision and instruction-following performance, often overlooking the fundamental aspect of creative design exploration. This study presents a TRIZ-inspired text-to-CAD framework that leverages LLMs to generate high-quality, editable CAD models while systematically exploring creative design alternatives. The framework integrates the Theory of Inventive Problem Solving (TRIZ)-embedding deep human insights from extensive patent records-into LLM prompting strategies, enabling autonomous generation of innovative CAD variants that address technical contradictions. Through a comprehensive three-stage pipeline of design generation, enhancement, and optimization, the framework produces structurally diverse CAD models from well-crafted prompts. The present study implements and evaluates the first two stages, while positioning the design optimization stage as future work. A product design case study (chair) demonstrates that the TRIZ-inspired text-to-CAD framework generates multiple creative design alternatives by systematically applying TRIZ inventive principles such as segmentation, anti-weight, dynamics, and composite materials, achieving 4.0-14.7% mass reduction across all enhanced designs while maintaining structural integrity. The key findings suggest that integrating systematic innovation methodologies with LLM-based 3D CAD generation bridges the gap between precision-focused synthesis and creativity-focused exploration, advancing toward autonomous design systems where AI makes design decisions independently, supporting human decision-making in human-AI collaborative design for engineering applications.
Dongeon Lee, Leekyo Jeong, Soyoung Yoo +2
Cho Chun Shik Graduate School of Mobility, KAIST, Daejeon, Republic of Korea · Samsung Electronics, Suwon, Republic of Korea · Department of Mechanical Engineering, Hanyang University, Ansan, Republic of Korea +1
AI-driven engineering workflows face particular challenges in crash safety design: unlike aerodynamics, crash events involve highly nonlinear contact dynamics, material nonlinearity, and discrete state transitions that are difficult to capture with data-driven surrogate models. To the best of our knowledge, we present the first foundation model--orchestrated workflow for crash safety design that enables surrogate-assisted exploration for pedestrian protection, reducing evaluation time from hours per CAE simulation to seconds. The workflow integrates four components: (1) a surrogate trained on CAE crash simulations to predict pedestrian leg injury metrics from design parameters, achieving an average R2=0.87 and providing distribution-free conformal prediction intervals; (2) multiobjective evolutionary search (NSGA-II) to discover diverse feasible parameter sets under user-specified constraints; (3) a morphing-based geometry generator that maps parameters to topology-preserving 3D shapes; and (4) a natural-language interface in which an LLM orchestrates the workflow and a vision--language model supports semantic comparison of generated designs. In an automotive front-bumper case study, the workflow produces 35 distinct safety-compliant alternatives from a single exploration, a process that would require weeks with conventional CAE iteration. These results suggest that foundation models can serve as integration layers between ML surrogates and physics-based simulation, helping bring AI capabilities to safety-critical engineering domains.