- 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
-
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.
-
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
-
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.
-
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 . Without runtime and memory scaling over different , 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 approaches the sample count.
Questions for authors
-
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.
-
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?
-
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
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
(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 simulations, which sample velocity in an unseen range from to km/s while keeping the oblique angle the same as the training set.
-
Smallplate - Angle: generated simulations, which sample the oblique angle in the unseen range from to while keeping the impact velocity the same as in the training set.
-
Smallplate - Resolution: generated simulations for each spatial resolution in 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.
| Condition | Model | ↓ MAE (m) | ↑ R² (%) | ↓ L₂² (%) |
|---|---|---|---|---|
| Overall | ||||
| Seen | A | () | 99.34 (± 0.41) | 0.31 (± 0.32) |
| B | () | 99.52 (± 0.35) | 0.17 (± 0.21) | |
| Unseen speed | A | () | 97.32 (± 1.58) | 2.29 (± 1.51) |
| B | () | 97.19 (± 1.85) | 2.43 (± 1.78) | |
| Unseen angle | A | () | 81.99 (± 12.32) | 4.81 (± 4.68) |
| B | () | 64.53 (± 35.74) | 9.10 (± 11.83) | |
| Projectile | ||||
| Seen | A | () | 97.97 (± 1.38) | 1.00 (± 0.73) |
| B | () | 98.55 (± 0.99) | 0.69 (± 0.51) | |
| Unseen speed | A | () | 89.84 (± 7.05) | 3.73 (± 2.30) |
| B | () | 90.77 (± 6.27) | 3.09 (± 1.88) | |
| Unseen angle | A | () | -20.45 (± 122.56) | 25.86 (± 12.70) |
| B | () | -56.08 (± 221.08) | 23.90 (± 12.88) | |
| Plate | ||||
| Seen | A | () | 99.30 (± 0.40) | 0.29 (± 0.30) |
| B | () | 99.47 (± 0.35) | 0.15 (± 0.20) | |
| Unseen speed | A | () | 97.26 (± 1.56) | 2.18 (± 1.45) |
| B | () | 96.99 (± 1.94) | 2.38 (± 1.79) | |
| Unseen angle | A | () | 87.94 (± 7.64) | 3.13 (± 3.17) |
| B | () | 68.22 (± 31.88) | 8.15 (± 11.48) |
(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 L2 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:
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.
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, therefore the basis can't be shared across simulation where Ω0 varies. However, it can be extended to families of geometries via a learned transport map t. We can reparameterize the state u defined on Ω0 to be some state g defined on an unit cube Ω□, where g=u∘t is a composition of u and transport map t. With this reparameterization, we can learn the basis on a shared domain Ω□ and the geometry of each simulation is described through a function t. 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 108 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 32 learnable basis functions + 4 affine basis, producing a total of 108 coefficients for each field (position and velocity) and body (projectile and plate). Thus, we have 4 separate PODs for each field and body, each with a $ 108$-dimensional spatial model. Similarly, to compare against FunctionalODE - B, we used 2 PODs, each for a body. For dynamic, we mirrored FunctionalODE's approach in both A and B variants.
Results POD consistently outperforms FunctionalODE across metrics.