Source-linked AI summary
KPI-Conditioned Generative Design of Automotive Hood Inner Panels: A Two-Stage Retrieval-Generation Pipeline with Surrogate-Based Performance Estimation
Sudeep Chavare
TL;DR
The paper addresses inverse hood-panel design, where KPI requirements do not uniquely determine discrete topology families. It proposes reachability-based family selection followed by conditional point-cloud generation and surrogate evaluation, finding that topology controls major variation but surrogate error can limit within-family discrimination.
Problem
Inverse design remains difficult because KPI requirements omit manufacturing and packaging information needed to select among discrete topology families.
Method
The pipeline retrieves structurally reachable topology families, generates point-cloud geometry within a selected family, and evaluates candidates with surrogate models.
Results
Topology dominates parametric variation by a factor of 3 to 7, while the generator improves on reference designs in 103 of 104 families and reaches 68% of the retrieval-oracle bound.
Takeaways & Limitations
Topology selection largely determines mass and stiffness, whereas stress offers more within-family freedom and is the main target for generation.
Takeaways & Limitations
The 4,096-point representation was discarded because its approximately 20 mm spacing measured sampling noise rather than shape, causing flat training and latent collapse.
Abstract
from arXiv · showhide
An inner hood panel must meet a deflection target, stay below a stress limit, and hit a mass target. Machine-learned surrogates have made the forward direction, geometry to performance, fast and routine. The inverse direction, producing geometry from a stated requirement, remains largely unaddressed for industrial parts whose design space is organized into discrete topology families rather than a continuous parameterization. This work presents a two-stage pipeline for that inverse problem. A reachability stage determines which topology families can satisfy a given requirement vector. A conditional variational autoencoder then generates point-cloud geometry within a selected family, and a neural-operator surrogate estimates the performance of each candidate. The pipeline is built entirely from public data and freely available compute, and is deployed as an interactive tool. The pipeline works, with qualifications that are reported as primary findings rather than caveats. The surrogate is accurate in aggregate, but its error is comparable to the performance differences it is asked to discriminate, which bounds what can be claimed for any individual generated design. That ratio of surrogate error to within-class signal is argued to be the quantity that determines whether a pipeline of this kind can work at all.
1. Introduction
The paper frames inverse hood-panel design as a topology-level problem: performance requirements do not contain the manufacturing and packaging information needed to select a family. It therefore combines reachability-based family retrieval with conditional generation and surrogate evaluation.
- 1. Introduction: The target is a manufacturable rib-and-cutout architecture satisfying deflection, stress, and mass requirements.The conventional workflow spends its costly effort selecting a carryover concept and constructing first geometry before simulation and iteration.
- 1. Introduction: Forward surrogates predict performance quickly, but obtaining geometry from specified performance remains substantially harder.The paper distinguishes routine geometry-to-performance prediction from the inverse task of performance-to-geometry generation.
- 1.1 Inverse design is underdetermined at the topology level: A feasible requirement triple corresponds to a mean of 3.4 distinct topology families, making the inverse mapping non-unique.The requirement vector is mapped to the space of manufacturable hood geometries, but its inverse is not uniquely determined.
- 1.1 Inverse design is underdetermined at the topology level: KPI-only conditioning cannot select the correct family because packaging, tooling, interfaces, and other topology determinants are absent from the inputs.The paper treats this as missing information rather than insufficient training.
- 1.2 Contributions: The pipeline retrieves structurally reachable topology families, then generates geometry within a selected family and estimates candidate performance.This split lets the model narrow the structural search space while leaving unobserved engineering constraints to the engineer.
- 1.2 Contributions: The work contributes quantified negative results, a deployed tool, and a generator that improves on reference designs in 103 of 104 families.The reported contributions include conditioning collapse, reconstruction-resolution limits, tangential dominance, and performance-estimation reporting limited to supported resolution.
2. Related Work
Prior work established engineering datasets, forward performance prediction, and property-guided generation, but this paper positions its contribution around learned generation for discrete, industry-validated hood topology families with explicit surrogate-resolution analysis.
- 2. Related Work: CarHoods10k supports representation learning, mechanical-performance prediction, and evolutionary topology optimization under manufacturability constraints.Its demonstrated uses include compact latent representations, metamodels, and optimization through decoded latent codes.
- 2. Related Work: AutoHood3D procedurally generates hood variants and uses machine learning for forward regression rather than learned performance-conditioned geometry generation.Its conditioning uses geometric descriptors, and the released text–point-cloud corpus does not include a trained performance-conditioned model.
- 2. Related Work: Metamaterial studies demonstrate property-guided point-cloud generation in latent spaces organized by unit-cell type.The paper identifies unit-cell type as an architectural parallel to topology family, while distinguishing its discrete industrial design setting.
3. Data and Geometry Representation
The study constructs a cleaned 9,479-design corpus across 104 hood topology families and represents geometries as fixed-size point clouds with stress-aware sampling and globally preserved scale. Data audits and sampling experiments expose important limits, including unusable low-density representations and restricted family generalization.
- 3. Data: The final corpus contains 9,479 usable designs across 104 families, after explicit identifier joins, failure filtering, and family-size filtering.The raw data contain separate CAD and finite-element failure modes, and the final corpus has a median of 95 variants per family.
- 3. Data: Geometric mass agrees with the manifest for an audited design but has 1.5% mean absolute error across 20 designs.The positive bias is consistent with shell-model thickness assumptions differing from solid-geometry volume calculations.
- 3.3 Point-cloud representation: Point clouds provide fixed-size, permutation-invariant geometry representations without relying on meshing connectivity or point correspondence.This avoids treating remeshing artifacts as design intent and supports Chamfer-based comparison.
- 3.3 Point-cloud representation: Sampling uses 30% uniform points and 70% curvature-weighted points, reduced by farthest-point sampling to 16,384 points.The curvature weighting targets fillets, rib terminations, and cutout edges where peak stress concentrates.
- 3.4 Normalization: A single global affine transform preserves relative scale because per-hood normalization would make mass unlearnable.The retained normalized diagonal coefficient of variation is 0.051, confirming non-zero size variation.
- 3.5 Sampling resolution: Representation density was selected by comparing between-design Chamfer signal against same-design subsampling noise.The 4,096-point configuration was discarded because approximately 20 mm spacing exceeded the geometric differences of interest, flattening training and collapsing the latent variable.
- 3. Data: The split tests unseen variants within known families, not generalization to unseen topologies.The identical split is shared by the surrogate and generator, preventing surrogate evaluation of generated designs trained directly on those designs.
4. Design Space Characterization
The design space is organized by discrete topology families, with topology dominating variation in mass and deflection while stress retains greater within-family flexibility.
- Mass is nearly locked by topology, with 1.9% within-family variation after family selection.
- Stress retains the greatest within-family freedom, whereas deflection is set by topology with modest trim.
- Deflection spans 5.21 to 5.38 mm, while stress spans 173.7 to 248.4 MPa across the same 98 designs.
- Rib and cutout architecture changes discretely between topology families rather than varying continuously.
- Family-level correlations are 0.506 for stress–deflection, 0.124 for deflection–mass, and −0.121 for stress–mass.
5. Methods
The method first retrieves topology families whose KPI envelopes can reach a requirement, then generates point-cloud geometry within a selected family and estimates its performance.
- 5.1 Stage one: reachability retrieval: A family qualifies when each KPI’s fractional shortfall is at most τ = 0.03, with tolerance measured relative to the target.
- 5.1 Stage one: reachability retrieval: Retrieval combines envelope shortfall with centrality because generation returns approximately the family mean rather than arbitrary envelope points.
- 5.2 Surrogate performance estimation: The surrogate performs global scalar regression from R^(N×3) to R³ using a Transolver trunk, pooling, and an MLP head.
- 5.3 Conditional generation: Within each family, the decoder deforms a fixed reference point cloud by predicting a displacement field conditioned on standardized KPI targets and a learned family embedding.
- 5.3 Conditional generation: A single shared model serves all 104 families because each family has approximately 21 training designs, while the shared model trains on 6,606 designs.
6. Results
Retrieval was calibrated for realistic targets, the surrogate tracked deflection and mass closely but had higher stress error, and generation improved geometry while showing conditional-mean compression.
- 6.1 Retrieval calibration: τ = 0.03 never starved on feasible realistic requests while filtering 104 families to approximately seven.
- 6.2 Surrogate performance: Increasing sampling density and dataset size reduced stress error by 19%, deflection error by 20%, and mass error by 26% relative to the initial configuration.
- 6.2 Surrogate performance: Deflection and mass track parity diagonals closely, while stress scatter widens above roughly 250 MPa where designs are sparse.
- 6.3 Generative model performance: The generator outperforms KPI-based retrieval by 1.7 mm and reaches 68% of the interval between the reference design and oracle bound.
- 6.3 Generative model performance: Across 104 families, 103 improve on their reference design, with a median improvement of 8.0% and median KPI sensitivity of 8.60 mm.
- 6.3 Generative model performance: Generation compresses extreme requests toward the family mean and covers approximately 34% of the requested span.
7. Negative Results
The negative results identify three limits: surrogate resolution can fail within families, surrogate-guided conditioning can collapse, and point-cloud reconstruction is constrained by sampling and displacement behavior.
- 7.1 Surrogate resolution limits per-design prediction: 29 families resolve no KPI, 33 resolve one, 31 resolve two, and 11 resolve all three, so aggregate accuracy does not ensure per-family discrimination.Across-family variance is three to seven times larger than within-family variance, while within-family stress variation spans 0.99% to 29.72%.
- 7.2 Surrogate-in-the-loop conditioning collapses: The auxiliary surrogate-prediction loss improved Chamfer distance to 0.0160, or 14.2 mm, but eliminated KPI conditioning across the requested range.Generation became identical to four decimal places at a family's minimum and maximum KPI values.
- 7.2 Surrogate-in-the-loop conditioning collapses: Surrogate-in-the-loop conditioning collapsed because the 7.34% stress error matched the 8.75% mean within-family spread and exceeded it in 57 of 104 families.The validation KPI loss plateaued near 0.10, consistent with predicting the conditional center rather than using the requested condition.
- 7.3 Mesh reconstruction is blocked by the generation objective: At 16,384 points, 4.7 mm mean nearest-neighbor spacing exceeds the 1.4 mm shell thickness and similarly sized fillet radii, limiting reconstruction of unsampled geometry.The reconstruction routes were evaluated on real point clouds, isolating the sampling-density limitation from generative error.
- 7.3 Mesh reconstruction is blocked by the generation objective: Generated displacement has essentially zero signed normal mean, but attributing this tangential dominance to Chamfer loss remains unresolved without measurements on real within-family pairs.The paper reports that tangential sliding may reduce Chamfer loss as effectively as normal deformation, while fixed-topology feature resizing can itself be in-plane.
- 7.3 Mesh reconstruction is blocked by the generation objective: The residual decoder preserves point cardinality, so it cannot add or delete material or introduce absent cutouts; its single-family depth change was 3.2 mm versus 6.5 mm for real members.The reported magnitude is an indication from one family and one requirement vector, not a characterized result.
- 7.4 Verified non-issues: Generated points remain on the template manifold, with median surface distance 3.8 mm and maximum 29.7 mm, ruling out free-space occupancy as the explanation.The raw per-index displacement statistic is treated as an artifact because independently sampled point clouds have no point correspondence.
8. Deployed Tool
The deployed tool combines requirement entry, family reachability, conditional generation, surrogate evaluation, and achieved-performance ranking. It reports resolvability and uncertainty with downloadable point-cloud outputs and three-dimensional candidate views.
- Pipeline workflow: The pipeline accepts stress, deflection, and mass targets plus priority weights, scores reachability across 104 families, and generates candidates from a shortlist.The interface bounds requirement sliders by the ranges present in the training corpus and lets users select how many ranked candidates to return.
- Ranking and evaluation: Candidates are ranked by achieved rather than envelope-predicted performance, demoting a family with a 22.7% stress error and retaining a result within 2.4%, inside surrogate error.Generating and scoring shortlisted candidates makes the ranking responsive to candidate-level estimates rather than family envelopes alone.
- Ranking and evaluation: Each KPI report includes requested values, surrogate estimates with error bands, signed deviations, family training ranges and means, and a resolvability verdict.The highest-ranked candidates are reported per family, with unresolved families presented at family level when individual discrimination is unsupported.
- Reporting and outputs: The interface states the surrogate error floor, marks deviations indistinguishable at that floor as “ok,” and provides generated point clouds in PLY format.The provenance footer accompanies every result.
- Reporting and outputs: Returned candidates include three-dimensional views and downloadable point-cloud outputs colored by distance from the family reference; solid mesh output is deferred.The deliverable is design intent for CAD reconstruction rather than manufacturing geometry.
9. Limitations and Future Work
The pipeline’s main limitations arise from weak normal-direction geometry changes, surrogate resolution relative to within-family variation, and evaluation restricted to known topology families. Future work therefore targets the training objective, surrogate accuracy and uncertainty, and broader topology coverage.
- Chamfer training permits tangential point redistribution, limiting normal surface change, geometric fidelity, and KPI conditioning strength.Proposed remedies include normal-direction supervision, spacing regularization, and an Earth Mover approximation.
- Reducing stress error below approximately 4% would make within-family estimates meaningful for most families rather than 45% of them.Higher-density sampling at fillets, field-level supervision, and richer geometric features are proposed options.
- High-fidelity finite element analysis remains the only instrument with sufficient resolution to verify KPI achievement for individual generated designs.A single-family study of approximately 20 designs could establish what the surrogate cannot verify.
- The reported error band is a global test-set statistic rather than a per-design estimate, motivating ensemble or Gaussian-process uncertainty quantification.Design-dependent uncertainty would sharpen the reachability test.
- The pipeline cannot generate topologies outside the 104 families in the dataset, and evaluation covers unseen variants of known families rather than unseen topologies.Latent interpolation between family embeddings is suggested, but validation would lack ground truth for nonexistent topologies.
10. Conclusions
The paper develops and deploys a two-stage system for retrieving feasible hood topology families and generating point-cloud geometries within them. It improves on family references and KPI-based retrieval, while showing that surrogate resolution and point-cloud training dynamics constrain individual-design claims.
- The deployed pipeline combines reachability-based topology retrieval, a conditional point-cloud autoencoder, and a Transolver performance surrogate.
- The generative model improves on the family reference design in 103 of 104 families, outperforms KPI-based retrieval, and reaches 68% of the retrieval-oracle bound.
- Surrogate error exceeds within-family KPI spread for most families, so per-design prediction is verifiable for only a minority.The tool reports resolvability by family and KPI rather than using uniform confidence.
- Chamfer-trained generation redistributes points tangentially at 6.2 times the normal deformation rate, with zero net normal motion.This finding is linked to modest geometric improvement, weak conditioning, and failed mesh reconstruction routes.
- Surrogate resolution should be compared with the within-class variation required by a task before validation or conditioning.
Appendix A. Training Failure Modes
Three training failures had distinct causes: noisy subsampled Chamfer measurements, dead decoder activations, and inconsistent random templates. Targeted corrections produced a usable generative model.
- Subsampling 2,048 of 8,192 points made self-comparison noise 0.0255, exceeding the real geometric signal of 0.008 by threefold.Full-cloud Chamfer reduced self-comparison to 0.00007 and corrected the flat-loss latent collapse.
- Unnormalized ReLU decoder activations became negative everywhere, producing zero inputs, zero gradients in most layers, and an unchanged template for 140 epochs.Group normalization with GELU activations resolved the failure.
- Random within-family templates made identical targets pair with varied starting geometries, so templates contributed variance rather than answer information.One fixed reference design per family aligned training with inference and resolved the issue.