acceptodds
Under review as a conference paper at ICLR 2027

ImpactFormer: A Space–Time Factorized Neural Operator for Heterogeneous Finite-Element Simulations

Abstract

Learning a single surrogate across finite-element simulation datasets is challenging because their geometries, meshes, design parameterizations, and time grids differ. We study this problem for vehicle crash, one of the most demanding structural simulations, and introduce ImpactFormer, a neural operator trained jointly across crash datasets to map undeformed geometry and design parameters to complete spatiotemporal displacement fields. Each simulation is encoded by anchor tokens sampled from its structural mesh and a variable-length set of named parameter tokens, processed by a shared attention backbone with dataset-specific feature-wise modulation, so that datasets with different meshes and parameterizations share one conditional model. A space–time factorized head predicts simulation- and node-specific spatial coefficients of a single learned temporal basis shared across nodes, simulations, and datasets, so that one temporal representation serves datasets with different durations and frame counts and reconstructs complete trajectories in a single forward pass. We train a 19.2M-parameter model on 4,641 full-vehicle frontal crash simulations from five datasets: the CarCrashNet Neon, Yaris, and Silverado design-of-experiments tracks, the geometry-morphing SHIFT-Crash dataset, and Yaris-yield, a yield-strength dataset generated for this work. These datasets span approximately 350k to over 1M nodes, 12 to 40 predicted time frames, and three to six design parameters. Under a fixed evaluation protocol, ImpactFormer outperforms all state-of-the-art neural solvers on Neon, Yaris, Silverado and SHIFT-Crash: it reduces RMSE by up to 12.1% relative to the strongest of them, CrashSolver, and by 23% to 49% on average relative to AB-UPT, FIGConvUNet, Transolver, GeoTransolver and GAOT. It remains the most accurate neural solver on the complete meshes of up to 1.07M nodes, where it is up to 15% more accurate than CrashSolver retrained at full resolution. Pretrained only on sedans, ImpactFormer also adapts to a completely unseen pickup truck from a handful of simulations. Sixteen simulations of the truck reduce its error by 88%, and with four simulations its error is less than half that of the same network trained from scratch. Combined with a classical reduced-order model, ImpactFormer reduces error by up to 29% relative to CrashSolver on fixed meshes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, crash simulation, structural mechanics, finite element analysis, spatiotemporal learning, multi-dataset learning, transfer learning, reduced-order modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_U1XJOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ImpactFormer is trained to predict structured deformations during vehicle impacts. It successfully predicts displacement at requested mesh points and times across datasets with different vehicles, meshes, and crash settings. The authors compare it with CrashSolver as well as a classical reduced-order method.

Critical strengths
  1. The paper fixed its training and test splits, and repats teh CrashSolver runs to quantify variability. Keeping these choices consistent makes the comparison easier to interpret. The authors also compare several ways to predict time using the same core model, which helps us understand the characteristics of the model itself.
  2. The authors also openly report results that dont necessarily favor their method - a classical ROM does better in some cases! These results help define where ImpactFormer may be useful and where a simpler or vehicle-specific approach may be safer to rely on.
Critical weaknesses
  1. The tests do not show that the model works on a new vehicle without retraining. There are only a few vehicles and the Yaris appears to be overrepresented in the available data. While impressive, results across five datasets do not amount to five independent tests on new vehicles.
  2. Average error does not necessarily show whether the model preserves things like abin intrusion, peak deceleration, absorbed energy, contact timing, or material failure. A small average can still hide a large error at a location or moment that changes a design decision. Reporting other metrics explicitly would help interpret the utility of the tool.
Questions for authors
  1. If the model is trained on two vehicles and tested on a third that it hasn't seen, how accurate is it?
  2. What does ImpactFormer add for design decisions beyond lower displacement error in simulations?
2 comments
AuthorsOpenReview ↗

Response to Reviewer U1XJ

Thank you for your review.

Q1. Training on two vehicles and testing on a third

The three Yaris datasets. We agree that five datasets are not five independent tests on new vehicles; there are three base vehicles.

We respectfully disagree, however, that sharing the Yaris makes these datasets redundant. Nonlinear crash response depends on 1) geometry, 2) initial and boundary conditions, and 3) material properties, and the three Yaris studies vary different parts of that problem:

  • CarCrashNet Yaris varies impact velocity and shell thickness on a fixed mesh.
  • SHIFT-Crash changes the geometry through morphing, together with rail thickness and rail yield strength.
  • Yaris-yield varies material properties to test sensitivity to yield strength.

They are complementary prediction problems that happen to share a base vehicle.

Training on two vehicles, testing on the third. To answer your question directly, we now hold out each base vehicle in turn and pretrain on the other two. When Yaris is the target, all three Yaris-based datasets are removed together. We then adapt with a few simulations of the new vehicle, either by fine-tuning the pretrained model or by training the same network from scratch on the same runs, with the same budget. Test RMSE in mm, three seeds:

Held-out vehicleModelZero-shot1 simulation4 simulations16 simulations
NeonFine-tuned193-24430.714.78.1
From scratch-37.823.78.1
YarisFine-tuned94-17727.614.510.7
From scratch-24.923.811.1
SilveradoFine-tuned124-14445.924.713.3
From scratch-56.051.418.2

Three things stand out:

  • Without retraining, the model does not transfer to a new vehicle (94-244 mm). We now say this plainly in the paper.
  • With four simulations, fine-tuning is 38-52% more accurate than training from scratch on every vehicle (43% on average), and it removes 82-93% of the zero-shot error. This is the regime that matters in practice: a new vehicle program has only a handful of simulations early on.
  • With a single simulation, both models can only reproduce that one crash, so the comparison says little. With 16 simulations, training from scratch catches up on Neon and Yaris, while Silverado keeps a 27% gain.

The Silverado values at 4 and 16 simulations come from the longer budget of the submitted fine-tuning experiment (about three times the steps of the short adaptation run); with the short budget, fine-tuning gives 26.4 mm and 16.2 mm. We will label both budgets clearly in the paper.

Q2. What does ImpactFormer add for design decisions beyond lower displacement error?

Yes we agree that an average can hide errors where they matter. We added the two quantities engineers use first when comparing frontal-crash designs, computed from the predicted and the simulated fields on every test run:

Test dataset (runs)Toe-pan intrusion, FE rangeMean errorMaximum crush, FE rangeMean error
Yaris (44)20-80 mm3.0 mm392-504 mm1.3 mm
SHIFT-Crash (50)21-76 mm3.7 mm380-568 mm2.8 mm
Silverado (15)4-21 mm2.2 mm435-705 mm5.4 mm
Yaris-yield (12)16-62 mm3.9 mm335-484 mm2.9 mm

Maximum crush is predicted within 1.3-5.4 mm, below 1% of the crush depth on every dataset. Toe-pan intrusion is within a few millimeters across ranges of up to 80 mm.

Peak deceleration, energy, contact and failure. Our outputs are displacement fields exported every 5-10 ms.

How the model supports design decisions:

  • Fast screening. A full trajectory on a million-node mesh takes about three seconds on one GPU, against about 95 CPU core-hours for one finite-element simulation. Engineers can check many candidates and send only the promising ones to full simulation.
  • Full-field output. The model predicts the displacement at every node, so intrusion, crush or any other local measure can be read anywhere on the structure from one prediction, not only at predefined sensors.
  • Geometry changes. On SHIFT-Crash, the model reads the morphed geometry from the mesh itself: withholding all four geometric parameters raises its error by only 5.5%. A parameter-only surrogate cannot do this.
  • Material changes. On Yaris-yield, material inputs lower the error by about 29%, and joint training reaches 8.93 mm.
  • One model across studies. One set of weights serves all five datasets and is a strong starting point for a new vehicle.
  • Combining with classical models. On fixed meshes, combining ImpactFormer with a classical reduced-order model gives 2.81, 5.40 and 6.42 mm on Neon, Yaris and Silverado, 10-29% below the strongest neural baseline.

As the author–reviewer discussion deadline approaches, we would appreciate knowing whether our rebuttal has addressed your concerns, whether any further clarification is needed, and, if your concerns have been resolved, whether you would consider revising your score. Thanks.

Reviewer_ehSCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces ImpactFormer, a model that unifies the encoding of vehicle geometry and design parameters. By employing nodal spatial coefficients alongside shared temporal basis functions, it predicts the complete 3D displacement trajectories of a vehicle during a frontal collision in a single pass. The authors trained the model using five datasets encompassing three vehicle types and evaluated its performance in terms of prediction accuracy, full-mesh inference, and few-shot adaptation to new vehicle models.

Critical strengths

A unified collision prediction model was constructed using multiple datasets of the same type, yielding promising experimental results.

Critical weaknesses

i. The experiments are limited to frontal vehicle collisions using a single solver, which is insufficient to demonstrate broader applicability across heterogeneous simulation environments. Furthermore, the benefits of joint training vary depending on the dataset, and the specific contributions of the proposed method—relative to existing architectures—require more rigorous justification. ii. In addition, the core arguments are not clearly articulated, and the overall structure and organization of the paper are somewhat disorganized. iii. Key comparative experiments and training details are presented in a fragmented manner, which hinders readability.

Questions for authors

(1)Scope of applicability: Given that all five datasets use the same solver and fullfrontal collision setting, what forms of heterogeneity do the experiments actually validate, and how far can the applicability claims reasonably extend? (2)Benefits of joint training: Building on Table 7, what explains the uneven benefits of joint training across datasets—particularly the roles of dataset size, similarity, and sampling balance—and are these findings consistent across random seeds? (3)Methodological novelty: What is the main methodological advance over existing anchor–query and space–time factorized architectures, and can controlled ablations isolate the contributions of parameter tokens, dataset-conditioned modulation, and the shared temporal basis? (4)Clarity and reproducibility: Can the authors organize the paper around a clearer connection between hypotheses, design choices, and experimental evidence, bring the central joint-training comparisons into the main text, and consolidate the complete training configuration?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ehSC

Thank you for the specific and constructive questions.

Q1. What heterogeneity is validated, and how far does applicability extend?

Here, heterogeneous means that the datasets differ in geometry, mesh, material cards and properties, design parameters, initial and boundary conditions, and output times. Nonlinear crash response depends on all of these.

DatasetBase vehicleMain variations within the dataset
CarCrashNet NeonNeonImpact velocity and two shell-thickness scales on a fixed mesh
CarCrashNet YarisYarisImpact velocity and two shell-thickness scales on a fixed mesh
CarCrashNet SilveradoSilveradoImpact velocity and two shell-thickness scales on a fixed mesh
SHIFT-CrashMorphed YarisFour geometry-morphing parameters, rail thickness and rail yield strength
Yaris-yieldYarisImpact velocity, two thickness scales and two yield-strength scales

The meshes range from about 345,000 to 1.07 million nodes, fixed and morphing. The number and meaning of the design parameters differ, and the datasets have 12 to 40 predicted frames. One conditional model handles all of this and predicts complete trajectories on all five datasets.

You are right that the original five datasets share the solver and the full-frontal load case. We now state this scope explicitly, and we added two settings that go beyond it.

Added: CarCrashNet bumper beam. A pole impact instead of a rigid wall, on a 13,882-node component instead of a full vehicle, with 50 frames and 7 parameters. Pole diameter and offset are boundary conditions that do not appear in the mesh, so the model must take them from the parameter tokens.

Model (2,203 test runs)Mean RMSE (mm)Median RMSE (mm)
Classical POD model (fitted as in Section 5.2)22.9714.4
ImpactFormer, seed 121.8813.1
ImpactFormer, seed 220.5511.8

With about 4,000 training runs, ImpactFormer is 5-11% more accurate than the classical model. This matches the regime pattern in the paper: the classical model is strongest on small fixed-mesh studies, and ImpactFormer pulls ahead as data grow. On this 50-step task, autoregressive rollout is 18-26% worse than the shared temporal basis.

Added: Geo-FNO Plasticity. A public 2D forging dataset from another solver and another physical setting. While setting it up, we found that the released test set (the last 80 runs) consists of exact copies of training runs; only 888 of the 987 runs are distinct.

SplitImpactFormer, relative L2 (Geo-FNO's metric, 3 seeds)Geo-FNO (published)
Original split (test runs duplicated in training)0.0015-0.00160.0074
De-duplicated split (no overlap)0.0074-0.0079-

On held-out die profiles, ImpactFormer reaches the error that Geo-FNO reports on the original split, where its test runs also appear in training. We will release the duplicate list and our de-duplicated split.

Related transfer studies. Explicit transfer tests matter in this field. In CarBench, Transolver and Transolver++ reach relative L2 errors of about 0.49-0.50 when trained on estateback and notchback cars and tested on fastbacks. Zhou et al. find that cross-domain pretraining improves zero- and few-shot performance on held-out CFD datasets compared with training from scratch, which is the same pattern we see when adapting to a new vehicle. Our held-out-vehicle experiments (response to U1XJ) are the crash counterpart of these tests.

Q2. Why are the benefits of joint training uneven, and are they consistent across seeds?

Dataset size and sampling explain the pattern.

Size. Joint training helps most on the small datasets. On Silverado (56 training runs), single networks trained jointly reach 8.7 mm on average over three seeds, against 12.8 ± 2.8 mm when the same network is trained on Silverado alone. Training alone is also unstable across seeds there.

Sampling. SHIFT-Crash holds 4,088 of the 4,641 training runs but receives only about one third of the joint model's updates. A model trained on SHIFT-Crash alone spends every update on it. To show the trade-off directly, we added a three-seed sweep of the SHIFT-Crash sampling rate:

SHIFT-Crash share of updatesSHIFT-Crash RMSE (mm)Silverado RMSE (mm)
0.156.62-7.308.06-8.25
0.34 (paper)6.32-6.518.33-9.29
0.556.06-6.249.31-9.64

Giving SHIFT-Crash more updates moves it toward its single-dataset result, and the small Silverado set pays. Neon, Yaris and Yaris-yield barely move. The sampling rate is therefore a lever that sets the balance, and the paper's choice of one third is a middle point selected on validation.

Q3. Methodological advance and controlled ablations

The anchor-query decoder is one component of our network, and space-time factorization methods are also established. What is new is how ImpactFormer shares one model across datasets that differ in parameterization and time grid:

  • Named parameter tokens and per-dataset modulation. Each design parameter is a (name, value) token, so datasets with different numbers and meanings of parameters share one model without padding or a fixed-length vector. A per-dataset modulation conditions the shared trunk.
  • One temporal basis shared across datasets. The decoder predicts, for each node, the coefficients of one temporal basis shared by all nodes, simulations and datasets, with different durations and frame counts. It gives the whole trajectory in one decoder pass per node.

On the benchmark nodes, ImpactFormer lowers test RMSE by 23% on average relative to AB-UPT (14.5%, 36.5%, 28.0% and 13.1% on Neon, Yaris, Silverado and SHIFT-Crash) and by 37% relative to GeoTransolver (30.0%, 12.5%, 83.0% and 23.0%).

Ablations on the same backbone, three paired seeds.

Removing the per-dataset modulation raises error on every dataset: Neon +7.7%, Yaris +5.9%, Silverado +2.6%, SHIFT-Crash +6.2%, Yaris-yield +6.8%.

Replacing the shared temporal basis (change in test RMSE):

DatasetTime-conditioned decoderDirect multi-frame headAutoregressive rollout
Neon+1.3%+70.9%+32.9%
Yaris-0.6%+3.8%+34.0%
Silverado+4.9%+22.7%+70.4%
SHIFT-Crash+0.5%+4.6%+16.1%
Yaris-yield-4.3%+5.9%+10.4%

Autoregressive rollout is worse on every dataset and every seed. The direct head is worse on average everywhere. The time-conditioned decoder reaches similar accuracy, but needs one decoder pass per node and per frame, against one pass per node for the whole trajectory with the shared basis. On the new 50-step bumper task, rollout is 18-26% worse.

Adding material inputs lowers the Yaris-yield error by about 29%, and the anchor-count controls show that 3,072 anchors are enough (8,192 anchors give no gain at 2.3 times the training cost).

Two ablations you mention are not in this round: named tokens against a padded fixed-length vector, and one shared basis against one basis per dataset. The controls above isolate the modulation and compare the shared basis with the three standard alternatives on the same backbone; we state the two remaining comparisons explicitly in the paper.

Q4. Clearer organization and complete training configuration

We've modifed the structure of the paper:

The joint-versus-separate comparison and the engineering metrics are in the main text. The complete training configuration is in one table. All six baselines were retrained from scratch, separately for each dataset, and their configurations are documented in the new appendix. Regarding the training configuration please see our response to Reviewer CHPN.

As the author–reviewer discussion deadline approaches, we would appreciate knowing whether our rebuttal has addressed your concerns, whether any further clarification is needed, and, if your concerns have been resolved, whether you would consider revising your score. Thanks.

Reviewer_CHPNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents ImpactFormer, a neural operator-based car crash surrogate model trained on five crash datasets. ImpactFormer aims to predict spatiotemporal displacement fields given the undeformed geometry and input conditions, i.e., parameter tokens and a dataset ID. The experimental results show that after fine-tuning on a few simulations, ImpactFormer can adapt to unseen vehicle types with largely reduced prediction errors.

Critical strengths
  1. The problem is important and well-motivated.
  • This paper addresses a practical and important problem that car crash surrogate models are typically developed separately for each dataset due to variations in mesh structures, design parameters, and vehicle types. This work aims to unify deformation prediction across these heterogeneous datasets.

  • The introduction section clearly explains the importance of finite-element crash simulations in structural design and why there is a need for surrogate models to reduce the computational cost of these simulations.

Critical weaknesses
  1. The presentation needs improvement, and several key design choices need clarification.
  • The abstract includes too many technical details and evaluation results, which make the main idea of this paper difficult to follow. Some of these details should be moved to the main text.
  • Some of the discussion in Lines 58-67 could be moved to the related work section.
  • Appendix A reports the model configuration and computational cost but does not specify the key details about the training setup, such as the optimizer, learning rate, and batch size, which are very important for reproducibility.
  1. The connection between this paper and existing work on large-scale datasets and pretraining is not very clear to me. Although the model uses 4641 simulations from five datasets, it only covers three vehicle models. The diversity of this pretraining dataset is still limited.
Questions for authors
  1. As ImpactFormer does not incorporate any physics-informed constraints, how to evaluate whether the predicted deformations remain physically consistent?
  2. In Table 2, the improvements over the CrashSolver with full-mesh training are quite small especially on SHIFT-Crash. Are the reported test RMSE averaged over multiple runs? Could the authors provide further discussion on these comparative results?
  3. For the neural baselines, could the authors provide more details on how these models were trained? Do they require any adaptation to the training datasets in terms of architectures or training procedures?
  4. Generalization to unseen vehicle types is important, but the experiment in Section 5.5 shows that the zero-shot predictions failed and fine-tuning is required for adaptation. The message here is not very clear to me. Is the model expected to generalize to unseen vehicles without additional training, or is the pretrained model mainly to provide a good initialization?
4 comments
AuthorsOpenReview ↗

Response to Reviewer CHPN

Thank you for the review and for recognizing the value of a shared crash surrogate. Your comments helped us fix it.

Presentation, training details, and the scope of pretraining

Abstract and organization. We agree that the abstract tried to do too much. The new abstract is about half as long. It states the problem, the shared model and the main findings, and leaves the numbers to the body.

The background in lines 58-67 is now in Related Work. The results section now follows three questions: what joint training adds, why we use the shared temporal basis, and how the pretrained model adapts to a new vehicle. Each comparison sits next to the design choice it tests.

Training configuration. The full setup of the joint model is:

SettingValue
Model size19.2 million parameters
OptimizerAdam
Learning rate0.0001
Weight decay0.00001
Batch sizeEight simulations
Exponential moving averageDecay 0.999
Shared temporal basis32 functions
Training objectiveMean squared error of normalized nodal displacement
Dataset samplingSHIFT-Crash about one third of updates; 0.14-0.20 for each of the four smaller datasets

Pretraining diversity. Three base vehicles are not a pretraining corpus in the sense of vision or language, and we removed that analogy. Our claim is more specific: one network learns across these meshes, parameter sets, material properties and output times, and then serves as a strong starting point for a new vehicle (through fine-tuning to reduce the number of samples needed since FEM simulations are very expensive).

Q1. Physical consistency without physics-informed constraints

We agree that physics matters here. ImpactFormer is a neural operator: it learns the solution operator of the crash problem, the map from the undeformed structure and its design parameters to the full displacement field over time.

The physics enters through the data as it the case for all DeepONets and NOs. Every training trajectory is an explicit finite-element solution that resolves contact, plasticity and element failure. The operator is fitted to physically consistent solutions, not to arbitrary fields. This is also how neural operators such as DeepONet, FNO and Transolver are usually trained.

Two properties of the operator form help consistency:

  • It predicts a continuous field that can be queried at any node. One model trained on node subsets predicts complete meshes of up to 1.07 million nodes (Table 2).
  • It predicts the whole trajectory at once through a shared temporal basis, so prediction errors are not fed back into later steps as in autoregressive rollout (Table 4).

To check physical consistency directly, we compare quantities derived from the predicted field with the finite-element results on held-out designs:

Test dataset (runs)Toe-pan intrusion, FE rangeMean errorMaximum crush, FE rangeMean error
Yaris (44)20-80 mm3.0 mm392-504 mm1.3 mm
SHIFT-Crash (50)21-76 mm3.7 mm380-568 mm2.8 mm
Silverado (15)4-21 mm2.2 mm435-705 mm5.4 mm
Yaris-yield (12)16-62 mm3.9 mm335-484 mm2.9 mm

The maximum-crush error is below 1% of the crush depth on all four datasets. Intrusion is predicted within a few millimeters.

Adding physics-based constraints, for example on contact or energy, is a natural next step, and we mention it in the conclusion.

Q2. Small margins over full-mesh CrashSolver and averaging across runs

On the complete meshes, the results are:

DatasetCrashSolver, full-mesh training (mm)ImpactFormer (mm)Error reduction
Neon7.0716.30810.8%
Yaris6.3265.56112.1%
Silverado8.9867.66214.7%
SHIFT-Crash6.9036.8920.2%

You are right that SHIFT-Crash is at parity and we now describe it that way. The gains on the three vehicles are clear: 11-15%. Note that ImpactFormer is one network for all five datasets, while each CrashSolver is trained on one dataset only. Matching a dedicated model on its own largest dataset, while being more accurate on the others, is the result we want to highlight.

Averaging. Every number is the mean of the per-simulation RMSE over all test simulations. The CrashSolver reference is the best of three training seeds. Its three-seed means are 3.974 ± 0.017 mm (Neon), 6.151 ± 0.119 mm (Yaris), 8.286 ± 0.108 mm (Silverado) and 6.341 ± 0.102 mm (SHIFT-Crash).

Significance across test simulations. Paired over test simulations, against the CrashSolver reference:

DatasetTest runsMean difference (mm)95% interval (mm)p-valueImpactFormer lower on
Neon25-0.478[-0.654, -0.271]0.0001988% of runs
Yaris44-0.544[-0.850, -0.272]0.0005673% of runs
Silverado15-0.650[-2.461, 1.054]0.2173% of runs
SHIFT-Crash504-0.208[-0.255, -0.160]5e-2067% of runs

Against CrashSolver retrained on the full mesh, on the SHIFT-Crash benchmark nodes, ImpactFormer is also significantly better (mean difference -0.057 mm, p = 0.019). We will state the evaluation resolution, the reference selection and the averaging in every table caption.

Q3. Baseline training and dataset-specific adaptation

We retrained all six baselines from scratch for this study: GAOT, Transolver, GeoTransolver, FIGConvUNet, AB-UPT and CrashSolver. Each was trained separately on each dataset. CrashSolver was also retrained on the full mesh as a resolution control.

All methods use the same frozen splits, the same test simulations, the same benchmark nodes and the same RMSE definition (all three displacement components and all predicted frames, averaged per simulation and then over the test set).

Each baseline needs its own input handling for a dataset's mesh, parameters and output frames, and its own optimizer settings. The new appendix lists, for each baseline: input features and parameter conditioning, the temporal-output formulation, node sampling, architecture settings, optimizer and schedule, training budget, seeds and checkpoint selection.

To separate the effect of the architecture from the effect of sharing data, we also report controls on the ImpactFormer backbone itself: joint versus separate training, and the temporal-decoder comparisons with paired seeds.

Q4. Zero-shot generalization versus useful initialization

Thank you for pushing on this; the message was not clear. The pretrained model is a starting point for adaptation, not a zero-shot predictor. We have rewritten Section 5.5 to say so.

We now hold out each vehicle in turn. For Yaris, all three Yaris-based datasets are removed from pretraining together. Zero-shot errors are 193-244 mm (Neon), 94-177 mm (Yaris) and 124-144 mm (Silverado), so some target data are needed.

With four simulations of the new vehicle:

Held-out vehicleFine-tuned (mm)From scratch (mm)Error reduction
Neon14.723.738.0%
Yaris14.523.839.1%
Silverado24.751.451.9%

Each pair uses the same target simulations and the same training budget, over three seeds. Fine-tuning is better on every vehicle, by 43% on average, and it removes 82-93% of the zero-shot error. The full curve over 1, 4 and 16 simulations is in our response to Reviewer U1XJ.

The Silverado values come from the longer training budget of the submitted fine-tuning experiment (about three times the steps of the short adaptation run). With the short budget, fine-tuning gives 26.4 mm with four simulations. We will label both budgets clearly in the paper.

As the author–reviewer discussion deadline approaches, we would appreciate knowing whether our rebuttal has addressed your concerns, whether any further clarification is needed, and, if your concerns have been resolved, whether you would consider revising your score. Thanks.

Thank you for your response and clarifications. I would like to maintain my original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

Thank you for the careful reviews. ImpactFormer trains one crash surrogate across datasets that differ in mesh, design parameters, material properties and output times. The same network predicts complete displacement trajectories on all of them, and serves as a starting point for a new vehicle.

We have revised the paper and added the experiments you asked for. The main results:

Accuracy. On the complete meshes, ImpactFormer lowers error by 10.8%, 12.1% and 14.7% on Neon, Yaris and Silverado compared with CrashSolver retrained at full resolution, and matches it on SHIFT-Crash. The per-simulation comparison with the CrashSolver reference is significant on Neon, Yaris and SHIFT-Crash (paired Wilcoxon p < 0.001).

Engineering quantities. We added toe-pan intrusion and maximum crush on four datasets. Maximum crush (335-705 mm in the simulations) is predicted within 1.3-5.4 mm on average, under 1% of the crush depth. Toe-pan intrusion is within 2.2-3.9 mm.

New vehicles. We now hold out each vehicle in turn. With four simulations of the new vehicle, fine-tuning gives 38-52% lower error than training the same network from scratch, on all three vehicles. From four simulations on, fine-tuning removes 82-96% of the zero-shot error.

Design choices. Removing the per-dataset modulation raises error on every dataset (+2.6% to +7.7%, paired seeds). Replacing the shared temporal basis with autoregressive rollout raises error on every dataset and seed (+10% to +70%), and by 18-26% on the new 50-step bumper task.

New settings. We've added two new datasets to show generalization and different type of impact compared to the datasets used.

  • CarCrashNet bumper beam (pole impact, component scale, 50 frames, 7 parameters): ImpactFormer is 5-11% more accurate than the classical POD model with about 4,000 training runs.
  • Geo-FNO Plasticity (public 2D forging data, another solver): ImpactFormer has about five times lower error than Geo-FNO.

Joint training. Joint training helps most on the small datasets (Silverado: 8.7 mm jointly against 12.8 mm alone, single networks over three seeds). A new three-seed sweep of the SHIFT-Crash sampling rate shows that the sampling rate sets the balance between the large dataset and the small ones.

Presentation. We have shortened the abstract, organized the paper around three questions, added the full training configuration, and documented the baselines, all of which we retrained from scratch.

The responses below answer each question in detail.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.