acceptodds
Under review as a conference paper at ICLR 2027

FunctionalODE: Reduced-Order Neural Dynamics for Extreme-Deformation Lagrangian Systems

Abstract

Lagrangian methods discretize a continuum into particles whose positions and properties are tracked over time. Learned graph-based simulators model these dynamics through repeated message passing over particle interaction graphs, so every particle is updated at every step, and the connectivity becomes unstable under extreme deformation and fragmentation. We propose FunctionalODE, a reduced-order alternative that views the system state as a vector-valued function over the fixed reference configuration. FunctionalODE learns an orthonormal basis for the low-dimensional subspace occupied by the states, encodes each state by projection onto this basis, and evolves the resulting compact coefficients with a NeuralODE, avoiding particle-level message passing; the full state can be evaluated at arbitrary material points. On SPH simulations of hypervelocity plate impact, FunctionalODE reaches on the fragmenting projectile, compared with for the best baseline, and completes a rollout in under a second, about faster than graph-based simulators and the numerical solver. FunctionalODE could serve as a cheap surrogate for expensive hypervelocity-impact simulations and enable the integration of simulation into prediction for physical impact experiments.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
hypervelocity impact, function encoders, neural ode, Lagrangian simulation

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_ca58OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents FunctionalODE, a reduced-order learned simulator for Lagrangian systems undergoing extreme deformation and fragmentation. Instead of evolving every particle through graph message passing, the method represents the state as a vector-valued function over a fixed reference configuration. A Function Encoder learns SIREN-parameterized orthonormal basis functions, each observed state is projected onto this basis through least squares, and a FiLM-conditioned Neural ODE evolves the resulting coefficients. An affine basis augmentation captures homogeneous deformation, and the full particle state can be decoded at arbitrary material coordinates. On a new SPH plate-impact dataset, FunctionalODE substantially improves projectile rollout accuracy over graph and transformer baselines while running in under one second. The paper also evaluates the method on 2D Navier-Stokes and identifies the fixed linear subspace as a limiting factor in that setting.

Critical strengths
  1. A well-motivated representation for fragmenting Lagrangian systems. Defining the state over fixed material coordinates avoids rebuilding interaction graphs as spatial neighborhoods change. The learned continuous basis, closed-form encoding, affine augmentation, and latent Neural ODE form a coherent and computationally attractive framework.

  2. Strong results on the plate-impact benchmark and transparent discussion of limitations. FunctionalODE achieves particularly large improvements on the fragmenting projectile and is much faster than graph-based simulators and the numerical solver. The authors also clearly report that the method underperforms at denser sampling on Navier-Stokes and correctly connect this behavior to truncation by the fixed linear subspace.

Critical weaknesses
  1. The experiments do not isolate the contribution beyond classical reduced-order modeling. FunctionalODE is closely related to a learned continuous POD-style basis followed by latent coefficient dynamics, yet the paper does not compare against POD/PCA with a Neural ODE or another reduced dynamics model. It also lacks architecture-matched ablations for the learned orthonormal basis, affine augmentation, FiLM conditioning, whole-rollout training, and Neural ODE. Given that the impact dataset has fixed geometry and material and varies mainly through impact speed and angle, its trajectory family may already occupy a very low-dimensional subspace. The current comparisons therefore do not establish that the proposed functional basis is responsible for the reported gains.

  2. The main generality and complexity claims are not sufficiently validated. The principal Lagrangian evaluation uses one geometry, one material setup, one particle resolution, and apparently in-range train/test conditions. There are no experiments on unseen geometries, materials, resolutions, or out-of-distribution impact parameters, even though the method requires a separate basis for each fixed reference configuration. Likewise, only latent integration is independent of particle count; full-state decoding costs O(NK)O(NK). Without runtime and memory scaling over different NN, the claim that rollout cost is decoupled from particle count is incomplete. Numerical stability is also a concern because encoding uses normal equations and reconstruction already degrades as KK approaches the sample count.

Questions for authors
  1. How does FunctionalODE compare with a classical POD/PCA basis followed by the same Neural ODE, and with an architecture-matched learned discrete basis? This comparison seems necessary to establish the value of the Function Encoder and its continuous basis.

  2. Can the authors evaluate generalization to unseen impact speeds or angles, different geometries or materials, and different particle resolutions? How is a new reference configuration handled without retraining its basis?

  3. What are the end-to-end runtime and peak-memory scaling curves as the number of particles increases, including full-state decoding at every reported time step? Could the authors also report the condition numbers of the least-squares systems and compare the normal-equation solve with a QR- or SVD-based implementation?

6 comments
AuthorsOpenReview ↗

(1/3) Reponse to question 1.

We thank the reviewer for the valuable comments, which point to weaknesses the submission did not address.

In the single-geometry setting, POD achieves lower reconstruction and rollout error than FunctionalODE (A1), and we will state this explicitly in the revised paper. This is expected: on a fixed discretization, the POD basis is the optimal linear basis in the L2L^2 sense. FunctionalODE gives up some of that optimality in exchange for a basis made of continuous functions on the reference domain, rather than vectors defined on one particular particle set. This difference has two consequences:

  1. POD is tied to a spatial resolution. POD modes are vectors defined only at the particles of the training discretization; using them at another resolution requires either recomputing the basis from new snapshots or interpolating the modes, an extra approximation outside the method. FunctionalODE's basis functions are defined at every point of the reference domain, so a trained model can encode and decode at a different resolution without retraining. Accuracy is not fully preserved across resolutions, since the basis cannot represent features finer than those seen in training; we quantify this in the resolution-transfer analysis at the end of this response.

  2. Long-term motivation: multiple geometries. FunctionalODE admitted a crittical limitation is that it only work on single reference configuration because it relies on an inner product that is an integral over Ω0\Omega_0, therefore the basis can't be shared across simulation where Ω0\Omega_0 varies. However, it can be extended to families of geometries via a learned transport map tt. We can reparameterize the state uu defined on Ω0\Omega_0 to be some state gg defined on an unit cube Ω□\Omega_\square, where g=u∘tg = u \circ t is a composition of uu and transport map tt. With this reparameterization, we can learn the basis on a shared domain Ω□\Omega_\square and the geometry of each simulation is described through a function tt. This extension is future work and not a claim of the current paper, but it is the setting the function-space design is built for.

We will revise the introduction and the limitations section accordingly, and we do not claim an accuracy advantage over POD on a fixed geometry.


Q1: How does a POD/SVD basis perform with the same setup?

A1:

  • Setup We implemented Random SVD with 108108 modes and compared it against our method. The justification for the number of SVD modes is to equate the number of coefficients in both methods. FunctionalODE - A has 3232 learnable basis functions + 44 affine basis, producing a total of 108108 coefficients for each field (position and velocity) and body (projectile and plate). Thus, we have 44 separate PODs for each field and body, each with a $ 108$-dimensional spatial model. Similarly, to compare against FunctionalODE - B, we used 22 PODs, each for a body. For dynamic, we mirrored FunctionalODE's approach in both AA and BB variants.

  • Results POD consistently outperforms FunctionalODE across metrics.

TaskFieldMetricUnitPODOur
ReconstructionxMAE×10⁻⁵2.93 (± 1.4)5.83 (± 2.34)
R²%99.85 (± 0.12)99.65 (± 0.25)
L²%1.99 (± 0.86)3.06 (± 1.13)
vMAE×10⁻⁵11.20 (± 3.31)17.78 (± 5.03)
R²%97.41 (± 1.50)94.85 (± 2.63)
L²%13.06 (± 3.76)18.55 (± 5.01)
Rollout AxMAE×10⁻⁵4.91 (± 2.45)16.89 (± 7.53)
R²%99.66 (± 0.34)99.45 (± 0.51)

Dear Reviewer,

As the rebuttal period is coming to an end, I would like to kindly ask you to read the authors' response and reply as soon as possible.

Thank you for your prompt attention.

Best regards,

Area Chair

AuthorsOpenReview ↗

(2/3) Response to question 2 - speed and angle

Q2: Can the authors evaluate generalization to unseen impact speeds or angles, different geometries or materials, and different particle resolutions? How is a new reference configuration handled without retraining its basis?

A2: In order to validate our method to unseen impact speeds, angles, and geometries, we generated three datasets as follows:

  • Smallplate - Velocity: generated 3232 simulations, which sample velocity in an unseen range from 5.05.0 to 7.57.5 km/s while keeping the oblique angle the same as the training set.

  • Smallplate - Angle: generated 3232 simulations, which sample the oblique angle in the unseen range from 30∘30^\circ to 60∘60^\circ while keeping the impact velocity the same as in the training set.

  • Smallplate - Resolution: generated 6464 simulations for each spatial resolution in {2×10−4,3×10−4,4×10−4}\{2\times 10^{-4}, 3\times 10^{-4}, 4\times 10^{-4}\} m, with the same impact velocity and oblique angle.

We trained our model on the original training dataset and evaluated on the new datasets without retraining the model. The following tables report evaluation metrics of variants A and B on each dataset.

Unseen speed and angle: FunctionalODE seems to generalize on out-of-distribution for impacting velocity and fails to generalize for impacting angle. The accuracy degradation is also more noticeable on variant B; this is to be expected since variant B directly use initial velocity vector as input for the dynamic ODE.

ConditionModel↓ MAE (m)↑ R² (%)↓ L₂² (%)
Overall
SeenA1.63×10−41.63 \times 10^{-4} (±7.96×10−5\pm 7.96 \times 10^{-5})99.34 (± 0.41)0.31 (± 0.32)
B1.01×10−41.01 \times 10^{-4} (±6.15×10−5\pm 6.15 \times 10^{-5})99.52 (± 0.35)0.17 (± 0.21)
Unseen speedA5.03×10−45.03 \times 10^{-4} (±1.89×10−4\pm 1.89 \times 10^{-4})97.32 (± 1.58)2.29 (± 1.51)
B5.15×10−45.15 \times 10^{-4} (±2.35×10−4\pm 2.35 \times 10^{-4})97.19 (± 1.85)2.43 (± 1.78)
Unseen angleA5.51×10−45.51 \times 10^{-4} (±3.13×10−4\pm 3.13 \times 10^{-4})81.99 (± 12.32)4.81 (± 4.68)
B7.25×10−47.25 \times 10^{-4} (±4.98×10−4\pm 4.98 \times 10^{-4})64.53 (± 35.74)9.10 (± 11.83)
Projectile
SeenA3.79×10−43.79 \times 10^{-4} (±1.82×10−4\pm 1.82 \times 10^{-4})97.97 (± 1.38)1.00 (± 0.73)
B3.05×10−43.05 \times 10^{-4} (±1.45×10−4\pm 1.45 \times 10^{-4})98.55 (± 0.99)0.69 (± 0.51)
Unseen speedA1.33×10−31.33 \times 10^{-3} (±4.95×10−4\pm 4.95 \times 10^{-4})89.84 (± 7.05)3.73 (± 2.30)
B1.11×10−31.11 \times 10^{-3} (±4.31×10−4\pm 4.31 \times 10^{-4})90.77 (± 6.27)3.09 (± 1.88)
Unseen angleA2.83×10−32.83 \times 10^{-3} (±1.74×10−3\pm 1.74 \times 10^{-3})-20.45 (± 122.56)25.86 (± 12.70)
B2.67×10−32.67 \times 10^{-3} (±1.62×10−3\pm 1.62 \times 10^{-3})-56.08 (± 221.08)23.90 (± 12.88)
Plate
SeenA1.58×10−41.58 \times 10^{-4} (±7.78×10−5\pm 7.78 \times 10^{-5})99.30 (± 0.40)0.29 (± 0.30)
B9.64×10−59.64 \times 10^{-5} (±5.99×10−5\pm 5.99 \times 10^{-5})99.47 (± 0.35)0.15 (± 0.20)
Unseen speedA4.86×10−44.86 \times 10^{-4} (±1.84×10−4\pm 1.84 \times 10^{-4})97.26 (± 1.56)2.18 (± 1.45)
B5.02×10−45.02 \times 10^{-4} (±2.32×10−4\pm 2.32 \times 10^{-4})96.99 (± 1.94)2.38 (± 1.79)
Unseen angleA5.02×10−45.02 \times 10^{-4} (±2.84×10−4\pm 2.84 \times 10^{-4})87.94 (± 7.64)3.13 (± 3.17)
B6.84×10−46.84 \times 10^{-4} (±4.74×10−4\pm 4.74 \times 10^{-4})68.22 (± 31.88)8.15 (± 11.48)
Reviewer_fEHXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

A reduced order simulation approach FunctionalODE combining Lagrange dynamics and NeuralODE integrator is proposed. It represents a deforming particle system using learned spatial basis functions.The coefficients of the basis evolve with a NeuralODE. The proposed approach avoids rebuilding particle interaction graphs during coefficient evolution. The main experiment uses a projectile and plate geometry at hypervelocity impact with varying ` speed and angle. The paper reports promising accuracy and fast rollout compared to known alternatives.

Critical strengths
  1. The state changes continuously without rebuilding interaction graphs.
  2. Strong metrics against the baselines
  3. The reported rollout is much faster than the graph models and numerical solver.
Critical weaknesses
  1. A POD/SVD basis with the same dynamics is an obvious choice for comparison and is missing. This comparison would clarify the benefit of the learned basis.
  2. The impact tests use one geometry and vary only speed and angle.
Questions for authors

Q1. How will a POD/SVD approach perform with the same setup? Q2. How R2 was calculated and averaged? Q3. Can you select a common prediction-only interval and report the accuracy and timings on it?

4 comments
AuthorsOpenReview ↗

(1/3) Response to question Q1 + Q2

We thank the reviewer for the valuable comments, which point to weaknesses the submission did not address.

In the single-geometry setting, POD achieves lower reconstruction and rollout error than FunctionalODE (A1), and we will state this explicitly in the revised paper. This is expected: on a fixed discretization, the POD basis is the optimal linear basis in the L2L^2 sense. FunctionalODE gives up some of that optimality in exchange for a basis made of continuous functions on the reference domain, rather than vectors defined on one particular particle set. This difference has two consequences:

  1. POD is tied to a spatial resolution. POD modes are vectors defined only at the particles of the training discretization; using them at another resolution requires either recomputing the basis from new snapshots or interpolating the modes, an extra approximation outside the method. FunctionalODE's basis functions are defined at every point of the reference domain, so a trained model can encode and decode at a different resolution without retraining. Accuracy is not fully preserved across resolutions, since the basis cannot represent features finer than those seen in training; we quantify this in the resolution-transfer analysis at the end of this response.

  2. Long-term motivation: multiple geometries. Because FunctionalODE represents the state as a function on a reference domain, it can be extended to families of geometries. Given a diffeomorphism t:Ω□→Ω0t:\Omega_\square \to \Omega_0 from the unit cube to the reference configuration, a state uu on Ω0\Omega_0 can be pulled back to g=u∘tg = u \circ t on Ω□\Omega_\square, and recovered as u=g∘t−1u = g \circ t^{-1}. The basis is then learned on the shared domain Ω□\Omega_\square, while the geometry is carried entirely by tt. POD could only share a basis across geometries if every geometry were first resampled onto a common grid on Ω□\Omega_\square; FunctionalODE works directly with each geometry's own particles. This extension is future work and not a claim of the current paper, but it is the setting the function-space design is built for.

We will revise the introduction and the limitations section accordingly, and we do not claim an accuracy advantage over POD on a fixed geometry.


Q1: How does a POD/SVD basis perform with the same setup?

A1:

  • Setup We implemented Random SVD with 108108 modes and compared it against our method. The justification for the number of SVD modes is to equate the number of coefficients in both methods. FunctionalODE - A has 3232 learnable basis functions + 44 affine basis, producing a total of 108108 coefficients for each field (position and velocity) and body (projectile and plate). Thus, we have 44 separate PODs for each field and body, each with a $ 108$-dimensional spatial model. Similarly, to compare against FunctionalODE - B, we used 22 PODs, each for a body. For dynamic, we mirrored FunctionalODE's approach in both AA and BB variants.

  • Results POD consistently outperforms FunctionalODE across metrics.

TaskFieldMetricUnitPODOur
ReconstructionxMAE×10⁻⁵2.93 (± 1.4)5.83 (± 2.34)
R²%99.85 (± 0.12)99.65 (± 0.25)
L²%1.99 (± 0.86)3.06 (± 1.13)
vMAE×10⁻⁵11.20 (± 3.31)17.78 (± 5.03)
R²%97.41 (± 1.50)94.85 (± 2.63)
L²%13.06 (± 3.76)18.55 (± 5.01)
Rollout AxMAE×10⁻⁵4.91 (± 2.45)16.89 (± 7.53)
R²%99.66 (± 0.34)99.45 (± 0.51)
L²%3.00 (± 1.41)5.12 (± 2.26)
vMAE×10⁻⁵15.22 (± 5.26)18.16 (± 5.93)
R²%94.66 (± 3.92)94.78 (± 3.51)
L²%18.45 (± 6.42)18.72 (± 6.28)
Rollout BxMAE×10⁻⁵4.80 (± 2.51)10.53 (± 5.80)
R²%99.67 (± 0.34)99.61 (± 0.39)
L²%2.95 (± 1.43)3.73 (± 1.88)

Table: Comparison between FunctionalODE and its classical counterpart


Q2: Which R$^2$ averaging convention was used for each method?

A2: The R$^2$ metric is averaged uniformly across the output dimensions.

AuthorsOpenReview ↗

(2/3) Reponse to question Q3

Q3: Can you report accuracy over a common prediction-only interval and clarify which operations are included in the timings?

A3:

  • To match the baseline methods' prediction window, FunctionalODE encodes the 6$^{th}$ time-step into coefficient space and uses NeuralODE to evolve the coefficient forward in time. The physical state is then recovered by decoding the coefficients and compared against the simulation’s ground truth. The result is reflected in the table below.
↓ MAE (m)↑ R² (%)↓ L₂² (%)
Overall
GNS3.62e-04 (± 6.44e-05)88.66 (± 4.31)2.49 (± 1.10)
SEGNN1.82e-04 (± 3.93e-05)95.23 (± 1.74)1.20 (± 0.64)
GNOT3.86e-04 (± 1.05e-04)73.72 (± 4.75)5.14 (± 2.54)
Galerkin Transformer1.14e-04 (± 3.77e-05)97.67 (± 1.05)0.51 (± 0.29)
ONO3.50e-04 (± 1.21e-04)74.97 (± 3.13)5.48 (± 3.16)
Transolver-V31.11e-04 (± 3.81e-05)97.64 (± 1.16)0.49 (± 0.29)
FunctionalODE (A)1.63e-04 (± 7.96e-05)99.34 (± 0.41)0.31 (± 0.32)
FunctionalODE (B)1.01e-04 (± 6.15e-05)99.52 (± 0.35)0.17 (± 0.21)
Projectile
GNS1.83e-03 (± 4.39e-04)48.20 (± 14.73)27.88 (± 9.88)
SEGNN1.21e-03 (± 3.26e-04)70.71 (± 13.18)15.51 (± 7.12)
GNOT3.60e-03 (± 9.62e-04)-22.01 (± 22.77)79.34 (± 13.86)
Galerkin Transformer7.21e-04 (± 1.94e-04)91.54 (± 4.68)4.55 (± 2.61)
ONO3.53e-03 (± 1.09e-03)-12.13 (± 8.83)72.20 (± 3.11)
Transolver-V37.16e-04 (± 2.18e-04)91.82 (± 4.81)4.52 (± 2.64)
FunctionalODE (A)3.79e-04 (± 1.82e-04)97.97 (± 1.38)1.00 (± 0.73)
FunctionalODE (B)3.05e-04 (± 1.45e-04)98.55 (± 0.99)0.69 (± 0.51)
Plate
GNS3.31e-04 (± 5.73e-05)89.28 (± 3.22)1.79 (± 0.88)
SEGNN1.60e-04 (± 3.51e-05)95.95 (± 1.38)0.78 (± 0.42)
GNOT3.18e-04 (± 8.74e-05)81.04 (± 1.77)2.93 (± 1.67)
Galerkin Transformer1.01e-04 (± 3.49e-05)97.66 (± 0.97)0.41 (± 0.25)
ONO2.83e-04 (± 1.01e-04)79.95 (± 4.71)3.36 (± 2.17)
Transolver-V39.78e-05 (± 3.47e-05)97.69 (± 0.99)0.39 (± 0.25)
FunctionalODE (A)1.58e-04 (± 7.78e-05)99.30 (± 0.40)0.29 (± 0.30)
FunctionalODE (B)9.64e-05 (± 5.99e-05)99.47 (± 0.35)0.15 (± 0.20)

Please note that GNS and SEGNN's relative L$^2$ are different from the number reported in Table 1. This is a reporting error made in the main paper's Table 1; while other methods (GNOT, Galerkin Transformer, ONO, Transolver, and FunctionalODE) report the squared L$^2$ relative error, GNS and SEGNN report the L$^2$ relative error. This correction doesn't change the relative performance between methods and will be reflected in the final version of the manuscript.

  • As for the timing, the operations included are encoding the first time step and decoding the rest of the trajectory, each over the full particle set.

Dear Reviewer,

As the rebuttal period is coming to an end, I would like to kindly ask you to read the authors' response and reply as soon as possible.

Thank you for your prompt attention.

Best regards,

Area Chair

AuthorsOpenReview ↗

(3/3) Additional Experiment on Spatial Resolution

Setup

  • Model. FunctionalODE Variant A (K=32K=32 basis functions, plain Neural ODE for the dynamics), trained only on the original Small plate data at particle spacing Δx=0.5\Delta x = 0.5 mm.
  • Test data. The same geometry re-simulated at Δx=0.4\Delta x = 0.4, 0.30.3 and 0.20.2 mm, with impact speeds and angles inside the training range. At Δx=0.2\Delta x = 0.2 mm the plate has about 16×16\times as many particles as in training (table below).
  • The trained model is applied unchanged: no retraining, no fine-tuning. The coefficients are fitted from a sampled subset of particles (as in training), the trajectory is rolled out in coefficient space, and the state is decoded and scored at every particle of the new discretization, with the training normalization statistics.
DatasetΔx\Delta x (mm)Δx\Delta x / train Δx\Delta xTest simulationsProjectile particlesPlate particles
Small plate0.51.0641366,400
Small plate_R40.40.86428015,000
Small plate_R30.30.66455226,136
Small plate_R20.20.4642,176100,000

Results. The model transfers to finer resolutions with a moderate loss of accuracy in position and a large loss in velocity:

  • Position stays accurate. Overall position R2R^2 goes from 99.2%99.2\% at the training resolution to 94.9%94.9\%, 94.3%94.3\% and 93.1%93.1\% at Δx=0.4\Delta x = 0.4, 0.30.3 and 0.20.2 mm, and the relative L22L^2_2 error stays below 2%2\%. The plate, which carries almost all of the particles, keeps R2≥92%R^2 \geq 92\%. The projectile degrades more, from 97.0%97.0\% to 78.3%78.3\%.

  • Velocity degrades substantially. Overall velocity R2R^2 falls from 94.3%94.3\% to 77.3%77.3\%, 74.1%74.1\% and 65.5%65.5\%. On the plate it falls from 91.3%91.3\% to 52.5%52.5\% at Δx=0.2\Delta x = 0.2 mm, so we do not consider the velocity prediction reliable at that resolution.

The basis was learned only from fields sampled at Δx=0.5\Delta x = 0.5 mm. A finer discretization resolves sharper, more localized structure near the impact zone that no training snapshot contained, and the velocity field, which is less smooth than the position field, is affected most. In short, the continuous basis lets a single trained model run at any resolution without retraining, but it does not by itself give resolution-independent accuracy.

PartFieldMetricSmall PlateSmall Plate R4Small Plate R3Small Plate R2
overallxMAE2.09×10−42.09 \times 10^{-4} (±\pm 1.24×10−41.24 \times 10^{-4})2.29×10−42.29 \times 10^{-4} (±\pm 1.17×10−41.17 \times 10^{-4})2.35×10−42.35 \times 10^{-4} (±\pm 1.19×10−41.19 \times 10^{-4})2.45×10−42.45 \times 10^{-4} (±\pm 1.22×10−41.22 \times 10^{-4})
R²99.23% (± 0.83%)94.88% (± 2.99%)94.25% (± 3.08%)93.10% (± 3.48%)
relative L22L^2_20.48% (± 0.63%)1.19% (± 0.96%)1.37% (± 1.05%)1.71% (± 1.24%)
vMAE1.96×1011.96 \times 10^{1} (±\pm 7.40×1007.40 \times 10^{0})3.10×1013.10 \times 10^{1} (±\pm 1.16×1011.16 \times 10^{1})3.26×1013.26 \times 10^{1} (±\pm 1.22×1011.22 \times 10^{1})3.52×1013.52 \times 10^{1} (±\pm 1.35×1011.35 \times 10^{1})
R²94.26% (± 3.87%)77.28% (± 3.59%)74.07% (± 4.53%)65.54% (± 4.52%)
relative L22L^2_24.29% (± 3.08%)21.18% (± 1.72%)23.10% (± 1.60%)28.30% (± 1.19%)
platexMAE2.03×10−42.03 \times 10^{-4} (±\pm 1.21×10−41.21 \times 10^{-4})2.19×10−42.19 \times 10^{-4} (±\pm 1.15×10−41.15 \times 10^{-4})2.21×10−42.21 \times 10^{-4} (±\pm 1.16×10−41.16 \times 10^{-4})2.32×10−42.32 \times 10^{-4} (±\pm 1.19×10−41.19 \times 10^{-4})
R²99.26% (± 0.84%)94.30% (± 4.25%)94.04% (± 4.43%)92.45% (± 5.37%)
relative L22L^2_20.45% (± 0.60%)1.07% (± 0.92%)1.14% (± 0.97%)1.51% (± 1.19%)
vMAE1.77×1011.77 \times 10^{1} (±\pm 6.65×1006.65 \times 10^{0})2.83×1012.83 \times 10^{1} (±\pm 1.09×1011.09 \times 10^{1})2.91×1012.91 \times 10^{1} (±\pm 1.13×1011.13 \times 10^{1})3.17×1013.17 \times 10^{1} (±\pm 1.27×1011.27 \times 10^{1})
R²91.25% (± 7.85%)68.01% (± 7.79%)66.59% (± 7.93%)52.45% (± 8.15%)
relative L22L^2_26.76% (± 6.12%)34.45% (± 7.14%)35.90% (± 7.03%)45.99% (± 5.73%)
projectilexMAE4.90×10−44.90 \times 10^{-4} (±\pm 2.87×10−42.87 \times 10^{-4})7.62×10−47.62 \times 10^{-4} (±\pm 2.75×10−42.75 \times 10^{-4})8.69×10−48.69 \times 10^{-4} (±\pm 2.95×10−42.95 \times 10^{-4})8.30×10−48.30 \times 10^{-4} (±\pm 2.88×10−42.88 \times 10^{-4})
R²96.97% (± 2.28%)87.25% (± 1.74%)80.17% (± 3.00%)78.27% (± 4.94%)
relative L22L^2_21.63% (± 1.33%)7.32% (± 1.29%)11.51% (± 1.90%)11.61% (± 3.15%)
vMAE1.08×1021.08 \times 10^{2} (±\pm 4.47×1014.47 \times 10^{1})1.75×1021.75 \times 10^{2} (±\pm 5.08×1015.08 \times 10^{1})1.99×1021.99 \times 10^{2} (±\pm 5.59×1015.59 \times 10^{1})1.94×1021.94 \times 10^{2} (±\pm 5.51×1015.51 \times 10^{1})
R²93.47% (± 4.48%)77.74% (± 4.19%)68.21% (± 6.04%)65.11% (± 8.31%)
relative L22L^2_22.38% (± 1.80%)8.22% (± 1.53%)12.10% (± 2.04%)12.23% (± 3.29%)
Reviewer_wGZwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FunctionalODE, a reduced-order framework for learning Lagrangian dynamics under extreme deformation and fragmentation. The method represents physical states using learned spatial basis functions defined on a fixed reference configuration, then evolves their coefficients with a Neural ODE. Experiments on hypervelocity impact simulations demonstrate improved prediction accuracy and substantially faster rollouts than graph-based simulators.

Critical strengths
  1. Well-motivated formulation. Representing states on a fixed reference configuration provides a natural way to handle deformation, separation, and fragmentation without rebuilding particle interaction graphs during rollout.

  2. Computational efficiency. Evolving compact coefficients reduces the cost of temporal integration and makes it independent of the particle count for a fixed basis size. The reported rollouts are substantially faster than graph-based simulators and the numerical solver, with speeds comparable to transformer-based methods.

  3. Strong prediction accuracy. FunctionalODE performs particularly well on the fragmenting projectile, improving over both graph-based and transformer-based baselines. The results suggest a useful balance between accuracy and computational cost for this challenging application.

  4. Trajectory-level dynamics training. The dynamics objective penalizes errors along integrated trajectories, directly addressing error accumulation during rollout. The reported long-horizon predictions provide encouraging evidence for this approach.

  5. Useful analysis and discussion of limitations. The paper examines reconstruction quality across basis sizes and evaluates an additional Eulerian problem. Its discussion of weaker performance on denser Navier–Stokes observations helps clarify the limits of the reduced representation.

Critical weaknesses
  1. Selection of the number of basis functions. Figure 2 examines reconstruction error for different basis sizes, but the rationale for choosing K=32K=32 remains unclear. Was this choice based on validation reconstruction error, rollout accuracy, or computational cost? Since accurate reconstruction does not necessarily imply accurate dynamics, a small ablation showing rollout accuracy and runtime across basis sizes would strengthen this choice.

  2. Physical relevance of the reconstruction objective. The least-squares encoding and squared reconstruction loss optimize average reconstruction accuracy, which may not adequately reflect localized fracture or fragmentation. Can the authors clarify whether the learned basis preserves physically important features that contribute relatively little to the overall error? Evaluation using a relevant physical metric, such as fragment statistics or conservation error, would help establish the physical fidelity of the predictions.

  3. Sufficiency of the reduced state for predicting dynamics. A representation that reconstructs snapshots accurately may still omit information needed to predict their future evolution. What evidence supports treating the retained coefficients as a sufficient state for the learned dynamics? In particular, variant B uses position coefficients and initial-velocity conditioning but does not explicitly evolve velocity. A discussion of when this representation is adequate, and when additional state or history may be necessary, would clarify the method's scope.

  4. Separate optimization of representation and dynamics. The basis is trained for snapshot reconstruction, while the Neural ODE is subsequently trained using a trajectory-level objective. This separation may discard features that have a small reconstruction contribution but a significant influence on future dynamics. Have the authors considered jointly fine-tuning the basis and dynamics using rollout loss? A limited comparison with the sequential procedure would help assess this trade-off.

Overall assessment. The paper presents a promising approach to efficient simulation of extreme deformation, supported by strong projectile prediction results. My main reservations concern whether the reconstruction-based representation preserves the information needed for physically faithful, long-term dynamics. Clarifying the basis selection and the consequences of separate representation and dynamics training would strengthen the contribution.

Questions for authors

Clarification requests:

  1. On what basis was K=32K=32 chosen — validation reconstruction error, rollout accuracy, or computational cost? A small ablation showing rollout accuracy and runtime across basis sizes would settle this.

  2. Does the learned basis preserve physically important features, such as localized fracture or fragmentation, that contribute relatively little to the overall reconstruction error? A physical metric such as fragment statistics or conservation error would help establish this.

  3. What evidence supports treating the retained coefficients as a sufficient state for the learned dynamics? In particular, variant B uses position coefficients and initial-velocity conditioning but does not explicitly evolve velocity. When is this representation adequate, and when might additional state or history be necessary?

  4. Have the authors considered jointly fine-tuning the basis and the dynamics using a rollout loss, rather than training them sequentially? A limited comparison with the sequential procedure would help assess the trade-off.

6 comments
AuthorsOpenReview ↗

(1/4) Response to question 1

Thank you for your constructive comments.


Q1: On what basis was K=32K=32 chosen — validation reconstruction error, rollout accuracy, or computational cost? A small ablation showing rollout accuracy and runtime across basis sizes would settle this.

A1: We chose KK by validation reconstruction error. The table below reports an ablation over KK:

  • Validation MSE (selection criterion): reconstruction mean-squared error on the validation set, evaluated at each time step on a sampled subset of particles.
  • Test reconstruction / Rollout A / Rollout B: mean absolute error (MAE) on the test set, evaluated at each time step on the full particle set. Rollout A and B refer to FunctionalODE variants A and B.
KKMSE (×10−8\times 10^{-8})MAE (×10−4\times 10^{-4})Rollout A (×10−4\times 10^{-4})Rollout B (×10−4\times 10^{-4})
415.990.971.511.74
87.360.711.460.88
124.870.631.490.81
164.280.601.480.87
32 (*)3.360.581.691.05
483.500.591.721.13
644.720.611.791.20

Table: Ablation study: validation's reconstruction loss (MSE), test reconstruction error (MAE), rollout reconstruction error (MAE) by the number of basis functions.

Reconstruction error decreases up to K=32K=32 and then increases slightly. Rollout error, however, is lowest at smaller bases (K=8K=8 for variant A, K=12K=12 for variant B) and grows for K≥32K \geq 32. A larger basis lowers the projection error but makes the coefficient dynamics harder to learn, and the rollout error reflects both. Selecting KK by validation reconstruction error therefore does not give the best rollout accuracy; selecting it by validation rollout error would have given a smaller KK and lower rollout errors than those reported in the paper. We will add this ablation and discussion to the appendix, and state explicitly in the main text that KK was chosen by reconstruction error.

AuthorsOpenReview ↗

(2/4) Response to question 2

Q2: Does the learned basis preserve physically important features, such as localized fracture or fragmentation, that contribute relatively little to the overall reconstruction error? A physical metric such as fragment statistics or conservation error would help establish this.

A2: Neither FunctionalODE nor POD tracks damage or fracture explicitly. Both represent the deformation as a continuous map from the reference configuration, so fragmentation appears only as particles separating in the predicted positions. "Fragments" are also not a quantity the simulation produces directly: they are identified in post-processing, and how many are found depends on chosen parameters (the linking length and the minimum fragment size). We therefore evaluate fragmentation in two complementary ways:

  • A2.1. The cumulative mass distribution of the projectile debris cloud;
  • A2.2. Fragment statistics, reported over a range of fragment-definition parameters.

Summary The learned basis preserves the large-scale outcome of the impact as well as POD does: the spatial spread of the debris cloud, the mass of the residual plate, and the count and mass of the larger fragments. It under-resolves the smallest fragments. This is consistent with a continuous representation that does not model fracture explicitly, and fragment counts at small NminN_\text{min} are also the statistics most sensitive to the post-processing parameters. We will add these analyses to the appendix, together with a per-simulation case study (fragment count over time, cumulative fragment-mass distribution, and sensitivity to the linking length), and list the loss of fine-scale fragmentation as a limitation.

A2.1: <u>Cumulative mass distribution:</u> cumulative mass distribution's profiles show agreement between ground-truth and surrogate models. This analysis, together with the mass distribution, will be included in the next revision. The mass distribution is estimated via KDE with varying kernel length ({2,3,4}×Δx\{2, 3, 4\} \times \Delta x).

Coordinate (mm)Ground truthPOD (A)POD (B)FunctionalODE (A)FunctionalODE (B)
xx-direction
-100.0%0.0%0.0%0.0%0.7%
-52.2%2.2%2.2%2.2%2.9%
012.5%12.5%14.7%15.4%12.5%
550.0%55.9%52.9%61.0%55.1%
1090.4%89.0%91.2%91.2%89.7%
1594.1%94.9%94.9%94.9%93.4%
2096.3%95.6%97.1%97.1%95.6%
2598.5%97.8%98.5%99.3%96.3%
3099.3%99.3%99.3%100.0%99.3%
yy-direction
-100.0%0.0%0.0%0.0%0.0%
-50.0%0.0%0.0%0.0%0.7%
01.5%5.9%3.7%0.7%4.4%
515.4%15.4%15.4%15.4%25.7%
1075.0%77.9%77.9%79.4%83.1%
1589.7%90.4%91.2%90.4%89.7%
2091.2%92.6%91.9%92.6%91.2%
2592.6%93.4%93.4%93.4%91.9%
3097.1%97.1%97.1%97.1%94.1%
3598.5%99.3%98.5%99.3%98.5%
40100.0%100.0%100.0%100.0%100.0%
zz-direction
-1514.0%15.4%17.6%6.6%12.5%
-1048.5%47.8%48.5%43.4%44.9%
-563.2%63.2%64.7%61.8%61.0%
076.5%77.2%75.7%76.5%76.5%
585.3%85.3%85.3%85.3%84.6%
1090.4%91.2%91.2%92.6%91.2%
1594.1%94.9%94.9%94.9%94.1%
2096.3%95.6%95.6%95.6%95.6%
2599.3%99.3%99.3%99.3%98.5%

Table: Comparison of cummulative mass distribution along x,yx,y, and zz direction between ground truth, POD and FunctionalODE.

A2.2: <u>Fragment statistic</u>: mixed results. While the largest fragment mass fraction analysis shows agreement between ground truth and surrogate models, the fragment count statistic shows disagreement between ground truth and surrogate models. The difference in fragment count shows up in both POD and FunctionalODE.

Specification: We perform two fragment statistics analyses, namely, largest mass fraction and fragment counts, with the following specifications:

  • We used the last frame of the simulation to compute the fragments statistic.
  • Fragment labels are computed by constructing a connectivity graph, where two neighboring particles are connected if the distance between two particles is less than 1.5Δx1.5 \Delta x.
  • A fragment is a connected component that has at least NminN_\text{min} particles; in the analysis, we choose Nmin∈{10,20,30}N_\text{min} \in \{10, 20, 30\}.
  • Any fragment with less than NminN_\text{min} particle is counted as isolated particles.
  • We set the particle's mass to 11 (mg) for the particle mass analysis, and we excluded the largest fragments from the mass analysis.
AuthorsOpenReview ↗

(2/4) Response to question 2 - cont

A2.2.1 - Fraction of mass of the largest fragment: The following table reports the fraction of mass of the residual plate relative to its original mass. The analysis shows agreement between ground truth, POD, and FunctionalODE.

Methodmean ± stdρ
Ground truth0.935 ± 0.028–
POD (A)0.931 ± 0.0300.96
POD (B)0.931 ± 0.0310.97
FunctionalODE (A)0.931 ± 0.0300.96
FunctionalODE (B)0.929 ± 0.0300.96

Table: Fraction of mass of the largest fragment is the residual plate; the fraction is computed over the mass of the plate.

</br> </br>

A2.2.2 - Fragment count: The following table reports the number of fragments per simulation. Both POD and FunctionalODE reported a lower number of fragments than the simulation. The gap between simulation and surrogate seems to attenuate as we increase the minimum particles per fragment. This suggests that the surrogate model's position is smoother than the ground truth.

NminN_\text{min}Methodmean ± stdρ
10Ground truth3.79 ± 1.48–
POD (A)2.02 ± 0.960.16
POD (B)2.02 ± 0.960.03
FunctionalODE (A)2.05 ± 1.080.22
FunctionalODE (B)2.19 ± 1.120.16
20Ground truth1.95 ± 0.94–
POD (A)1.37 ± 0.600.49
POD (B)1.25 ± 0.440.56
FunctionalODE (A)1.35 ± 0.540.50
FunctionalODE (B)1.29 ± 0.490.47
30Ground truth1.44 ± 0.59–
POD (A)1.27 ± 0.510.64
POD (B)1.25 ± 0.440.67
FunctionalODE (A)1.24 ± 0.470.59
FunctionalODE (B)1.25 ± 0.470.65

Table: Fragment count comparison between the simulation's ground truth, POD, and FunctionalODE.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.