acceptodds
Under review as a conference paper at ICLR 2027

CoLEP: Co-Evolving Particles and Persistent Eulerian Latent Fields for Neural Fluid Simulation

Abstract

Graph neural networks (GNNs) have emerged as a powerful framework for learning particle-based fluid dynamics. However, GNN-based simulators suffer from rapid error accumulation over long rollouts and the high computational cost of repeated message passing. We introduce CoLEP, a co-evolving Lagrangian–Eulerian particle simulator that augments particle-carried states with a persistent latent field on a fixed Eulerian grid. At each step, current particle information updates the inherited Eulerian state, which is then fed back to guide particle prediction, allowing spatial interaction context to accumulate and be reused across time. Building on this persistent state, we introduce periodic GNN skipping (PGS), which interleaves full graph-based interaction updates with learned graph-free Fast steps while continuing to evolve both the particle and Eulerian states. Across four particle-simulation benchmarks, CoLEP consistently improves long-horizon rollout accuracy over the evaluated baselines. Eulerian state ablations further show that persistence is particularly important when graph evaluations are sparse, while PGS provides a favorable accuracy–inference-cost trade-off.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
particle-based simulation, neural surrogate modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ypHCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The proposed methods tackles two problems that occur in particle fluide simulations with neural methods:

  • error accumulations of rollouts and
  • high computational cost of message passing if graph neural networks are used. Their main trick is to keep a grid of hidden/latent features between timesteps. The model then moves information from particles to the grid, updates the grid, then moves information back to the particles. I utilizes a graph network before and after the grid step. A second trick (PGS), skips the graph networks on some steps, via the help of some smaller networks.
Critical strengths

The presented results are strong. Table 1 shows that CoLEP is beating the baselines on all of the rollout metrics across the four datasets.

Critical weaknesses

I am missing a discussion/comparison section for other latent-space methods. There have been other methods that are fairly similar, to the proposed method. UPT (Alkin et al., 2024) encodes physical data, runs dynamics in latent space, and decodes predictions. GINO (Li et al., 2023) moves irregular-point data onto a regular latent grid, processes it, and maps it back. MultiScale MeshGraphNets also uses coarse representations to make learned simulation cheaper. CoLEP is not the same: its fixed grid remembers information while particles keep moving, and the two update each other at every step. But the paper should explain this difference much more clearly. UPT (or the followup AB-UPT (Alkin et al., 2025)) is worth considering as a baseline if it can be adapted to the same particle rollouts.

Questions for authors
  1. Table 1, how many training seeds were used for each method and dataset? Can you show variation or uncertainty for the full-rollout scores, not just the error curves?
  2. Can you give a side-by-side table of parameter counts, training steps, training time, and inference setup for CoLEP and each baseline? Appendix B.5 gives some training details, but it is hard to judge whether the budgets are matched.
  3. Why is the repeated particle-grid update better than rolling forward in a latent space as UPT does? Can you compare against UPT on a matching particle task?
  4. Can you compare PGS against a smaller or shallower full-step model at about the same inference time? This would test whether skipping graph steps is better than simply using a cheaper model.
6 comments

Thanks for the careful assessment and constructive feedback. We appreciate your recognition of CoLEP's strong experimental results across all four benchmarks.

We would like to reiterate our two main contributions: (1) a persistent Eulerian latent field that co-evolves with Lagrangian particles, enabling spatial information to be accumulated and reused across time steps to improve long-horizon prediction; and (2) Periodic GNN Skipping (PGS), which leverages this persistent field to reduce expensive particle interactions while maintaining prediction accuracy.

We address your concerns and provide additional experimental evidence below.

Response to W1: Latent-Space Method Comparison

Thanks for your constructive suggestion. We fully agree that expanding the related work section makes our literature review more comprehensive and that clarifying the differences between CoLEP and UPT [1] is necessary. To address this, we revised the paper in two aspects. First, we expanded the Related Work section in the main text to introduce UPT [1], GINO [2], and MultiScale MeshGraphNets [3]. Second, we added a comparison section in Appendix F to discuss the architectural differences between CoLEP and related latent-space methods, along with experimental comparisons against a representative method of latent method UPT.

Regarding whether UPT is a suitable baseline for this task, we examine this question from both architectural design and experimental validation. First, from an architectural perspective, alongside dedicated particle-based simulators such as GNS and CoRGI, our CoLEP focuses on Lagrangian particle simulation and requires accurately predicting individual particle trajectories over long autoregressive rollouts. In contrast, UPT is primarily designed as a general-purpose framework for continuous field modeling and velocity field prediction. Because of this fundamental difference in task formulation, UPT's architecture is not tailored to capture particle interactions that are essential for preventing error accumulation in particle simulation tasks.

In Response to Q3, we provide a detailed architectural analysis explaining how these design choices affect its behavior on pure particle simulation tasks. To further validate this empirically, we evaluated UPT on several benchmark datasets, and the results confirm that while it can be adapted to particle rollouts, it is less competitive than dedicated neural particle simulators such as GNS and CoRGI. Detailed experimental settings, quantitative results, and further analysis are provided in Response to Q3 and have been updated in Appendix F.

References

[1] Benedikt Alkin, Andreas Fürst, Simon Schmid, Lukas Gruber, Markus Holzleitner, and Johannes Brandstetter. Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators. Advances in Neural Information Processing Systems, 37, 2024.

[2] Zongyi Li, Nikola Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Otta, Mohammad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, and Anima Anandkumar. Geometry-Informed Neural Operator for Large-Scale 3D PDEs. Advances in Neural Information Processing Systems, 36, 2023.

[3] Meire Fortunato, Tobias Pfaff, Peter Wirnsberger, Alexander Pritzel, and Peter Battaglia. MultiScale MeshGraphNets. 2nd AI4Science Workshop at ICML, 2022.

Responses to Questions

Response to Q1: Multi-Seed Benchmark Results

Thanks for pointing this out. In the initial submission, the results reported in Table 1 were obtained using a single training seed. To properly quantify variation and uncertainty, we have retrained and evaluated all methods across 3 independent random seeds for each dataset in the revised manuscript.

Specifically, we have added the complete 3-seed results in Appendix D of the revised manuscript, reporting the mean and standard deviation for every metric, and we also provide the full table below for convenience.

DatasetModelMSE₂₀ ↓KE–MSE₂₀ ↓MSE_full ↓KE–MSE_full ↓Sinkhorn_full ↓
DAM-2DGNS(4.47 ± 0.66) × 10⁻⁵(3.29 ± 0.97) × 10⁻⁵(3.17 ± 2.57) × 10⁻¹(1.39 ± 1.29) × 10⁻²(4.34 ± 3.19) × 10⁻²
NeuralMPM(5.27 ± 0.70) × 10⁻⁵(3.53 ± 0.31) × 10⁻⁴(2.82 ± 0.40) × 10⁻²(5.66 ± 1.84) × 10⁻⁴(9.48 ± 6.15) × 10⁻⁴
CoRGI(1.76 ± 0.23) × 10⁻⁵(2.14 ± 1.96) × 10⁻⁵(8.26 ± 5.27) × 10⁻²(2.21 ± 1.99) × 10⁻³(4.83 ± 4.93) × 10⁻³
CoLEP (Ours)(1.31 ± 0.01) × 10⁻⁵(2.66 ± 0.15) × 10⁻⁶(2.16 ± 0.10) × 10⁻²(3.35 ± 0.78) × 10⁻⁵(2.52 ± 0.26) × 10⁻⁴
TGV-2DGNS(7.04 ± 0.28) × 10⁻⁶(2.40 ± 0.26) × 10⁻⁷(5.74 ± 0.04) × 10⁻⁴(2.23 ± 0.23) × 10⁻⁷(9.85 ± 2.56) × 10⁻⁷
NeuralMPM(2.37 ± 0.04) × 10⁻⁵(6.01 ± 2.21) × 10⁻⁶(9.59 ± 0.41) × 10⁻⁴(4.13 ± 1.55) × 10⁻⁶(8.99 ± 1.18) × 10⁻⁶
CoRGI(5.91 ± 0.33) × 10⁻⁶(1.44 ± 0.30) × 10⁻⁷(4.81 ± 0.21) × 10⁻⁴(1.39 ± 0.58) × 10⁻⁷(1.14 ± 1.05) × 10⁻⁶
CoLEP (Ours)(3.55 ± 0.07) × 10⁻⁶(5.96 ± 0.85) × 10⁻⁸(3.23 ± 0.06) × 10⁻⁴(6.09 ± 1.23) × 10⁻⁸(2.04 ± 0.82) × 10⁻⁷
Moving Obstacle-2DGNS(9.78 ± 0.62) × 10⁻⁶(7.24 ± 0.89) × 10⁻⁷(2.03 ± 1.15) × 10⁻²(2.84 ± 0.79) × 10⁻⁵(2.26 ± 1.70) × 10⁻²
NeuralMPM(2.84 ± 0.21) × 10⁻⁵(1.38 ± 0.35) × 10⁻⁶(1.70 ± 0.10) × 10⁻³(2.00 ± 0.24) × 10⁻⁶(9.03 ± 0.46) × 10⁻⁴
CoRGI(6.68 ± 0.19) × 10⁻⁶(4.07 ± 1.02) × 10⁻⁷(3.39 ± 0.26) × 10⁻³(4.23 ± 0.35) × 10⁻⁶(2.39 ± 0.51) × 10⁻³
CoLEP (Ours)(1.61 ± 0.06) × 10⁻⁶(1.52 ± 0.01) × 10⁻⁷(8.92 ± 0.96) × 10⁻⁴(5.58 ± 1.13) × 10⁻⁷(2.47 ± 0.08) × 10⁻⁴
Water-3DGNS(5.63 ± 0.41) × 10⁻⁵(4.97 ± 2.30) × 10⁻⁴(8.75 ± 2.29) × 10⁻²(9.66 ± 6.52) × 10⁻³(1.10 ± 0.75) × 10⁻¹
CoRGI(4.00 ± 0.30) × 10⁻⁵(3.28 ± 0.13) × 10⁻⁴(6.79 ± 1.92) × 10⁻²(3.27 ± 1.88) × 10⁻³(6.53 ± 5.56) × 10⁻²
CoLEP (Ours)(3.72 ± 0.15) × 10⁻⁶(1.91 ± 0.68) × 10⁻⁵(1.42 ± 0.02) × 10⁻²(3.87 ± 0.69) × 10⁻⁵(6.28 ± 2.65) × 10⁻³

Table R1-1: Quantitative comparison of 20-step and full-rollout performance across all benchmark datasets. Results are reported as mean ± standard deviation across 3 independent random seeds, with the best scores highlighted in bold.

Response to Q2: Model Costs and Training Protocols

To clarify the computational budgets and evaluation protocols across methods, we report the exact parameter counts across datasets in Table R2-1, training budgets and wall-clock training times on DAM-2D in Table R2-2, as well as CoLEP's training schedules and grid configurations in Tables R2-3 and R2-4, followed by a summary of our unified inference setup.

Table R2-1 reports the exact parameter counts for CoLEP and each baseline across the benchmark datasets. Note that NeuralMPM is restricted to 2D domains and is therefore omitted on Water-3D.

DatasetCoLEP FullGNSNeuralMPMCoRGI
DAM-2D3,407,3381,211,7941,117,20017,437,586
TGV-2D3,966,1781,211,5381,117,20017,437,330
MovingObstacle-2D3,373,5381,212,0501,116,4184,369,170
Water-3D8,481,7631,213,203—11,318,451

Table R2-1: Exact parameter counts across benchmark datasets.

Because our training runs were distributed across NVIDIA A100 and RTX 4090 GPUs, wall-clock training times cannot be directly aligned across all datasets. Table R2-2 therefore only reports the training times on DAM-2D, where all four methods were trained on the same NVIDIA A100 GPU. Across all experiments, we followed the official implementations and standard training protocols for each baseline so that all models reached full convergence. Specifically, we used the LagrangeBench implementation of GNS and trained each seed for 500,000 optimizer updates across all datasets. NeuralMPM and CoRGI were trained using their official rollout protocols, using 101,000 updates with a batch size of 128 for NeuralMPM, and 500,000 updates with a batch size of 4 for CoRGI. CoLEP employs multi-step rollout curriculum supervision that progressively scales from 2 to 20 steps, requiring 30,000 optimizer updates on DAM-2D while achieving a wall-clock training time comparable to CoRGI.

MethodBatch SizeTraining Steps per SeedTraining Time (h)
CoLEP430,00028.99
GNS1500,0007.11
NeuralMPM128101,00019.86
CoRGI4500,00027.12

Table R2-2: Training budgets and wall-clock training time per seed on DAM-2D.

For completeness, Table R2-3 details CoLEP's total training steps and progressive rollout transition ranges across all five datasets, and Table R2-4 lists the corresponding core and computational grid resolutions along with particle interaction radii.

For inference setup, all methods are evaluated on the same single A100-SXM4-40GB GPU with a batch size of 1. GNS, CoRGI, and our CoLEP variants perform single-step autoregressive rollouts initialized from 6 historical frames, whereas NeuralMPM predicts 8 frames per forward pass normalized over the 20-step horizon.

DatasetTotal training steps per seedR2 → R4 transitionR4 → R8 transitionR8 → R16 transitionR16 → R20 transitionFixed R20 training range
DAM-2D30,0001,001–1,5002,751–3,5005,251–6,2508,501–10,00010,001–30,000
TGV-2D50,0002,779–4,1677,640–9,72214,584–17,36123,612–27,77827,779–50,000
MovingObstacle-2D18,0001,001–1,5002,751–3,5005,251–6,2508,501–10,00010,001–18,000
Water-3D39,50010,001–15,00024,251–25,00026,751–27,75030,001–31,50031,501–39,500
TGV-3D20,0001,001–1,5002,751–3,5005,251–6,2508,501–10,00010,001–20,000

Table R2-3: CoLEP training steps and progressive rollout curriculum schedules across datasets.

DatasetCore gridComputational gridParticle interaction radius
DAM-2D64 × 16072 × 1680.029
TGV-2D64 × 6464 × 640.029
MovingObstacle-2D64 × 6472 × 720.015
Water-3D32 × 32 × 3236 × 36 × 360.035
TGV-3D32 × 32 × 3232 × 32 × 320.46

Table R2-4: CoLEP grid resolutions and particle interaction radii across datasets.

Reviewer_Q8reOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CoLEP, a hybrid Lagrangian-Eulerian neural simulator that augments particle-carried states with a persistent latent field on a fixed Eulerian grid. At each rollout step, particle features and Pre-GNN interaction residuals are scattered to the grid, combined with the inherited grid state, and evolved by a recurrent U-Net. The resulting field is gathered back to the particles and refined by a Post-GNN before acceleration prediction and integration. The paper additionally proposes periodic GNN skipping (PGS), which alternates complete graph-based transitions with graph-free Fast transitions implemented through lightweight adapters while retaining the Eulerian state. On DAM-2D, TGV-2D, MovingObstacle-2D, and Water-3D, CoLEP reports lower long-horizon position, kinetic-energy, and particle-distribution errors than the evaluated baselines. Memory ablations and interventions indicate that carrying the grid state across rollout steps improves prediction, particularly when GNN evaluations are skipped.

Critical strengths
  1. Turning the Eulerian representation from a transient computational pathway into an evolving latent state is a clear and potentially useful extension of existing hybrid particle-grid simulators. The matched memoryless ablations provide meaningful evidence that cross-step persistence contributes beyond per-step grid processing.

  2. CoLEP improves all reported full-rollout metrics across four benchmarks, and the reset, frozen-state, and mismatched-state interventions provide useful insight into the trajectory-dependent information retained by the Eulerian field. PGS is also a practically relevant attempt to reuse this information to reduce repeated graph computation.

Critical weaknesses
  1. CoLEP is evaluated only within dataset-specific materials, physical parameters, particle resolutions, and fixed Eulerian grids, so it is unclear whether the persistent field generalizes across material and viscosity regimes or remains efficient as particle count and grid resolution grow. Can the model interpolate or extrapolate across viscosity values and transfer to unseen materials, and how do parameter count, grid size, peak memory, training and inference time, and rollout error scale with particle count? Please add held-out material and viscosity experiments together with controlled particle-count and grid-resolution scaling on matched hardware, including a component-level runtime analysis and 3D PGS costs.

  2. Three of the four benchmarks are 2D, while Water-3D is the only evaluated 3D regime; persistence and PGS are also not validated in 3D. Does the reported advantage extend to the granular dynamics of Sand-3D and the highly viscous or non-Newtonian dynamics of Goop-3D? Please evaluate both datasets; fixed representative subsets are acceptable if the full datasets are prohibitively expensive. The comparison should include CoLEP, its no-persistence variant, and relevant baselines, while clearly reporting subset construction, train/validation/test splits, training budgets, checkpoint-selection rules, seed counts, and mean and standard deviation across runs.

If these experiments demonstrate robust cross-material and viscosity generalization together with favorable 3D scaling under a transparent multi-seed protocol, I would raise my rating to 3.

Questions for authors
  1. How does CoLEP generalize across materials and physical parameters? In particular, can the authors report held-out viscosity interpolation and extrapolation, as well as joint training or transfer across water, sand, and goop? How would the current categorical material embedding represent a material or parameter regime not observed during training?

  2. As particle count increases, how do model parameters, grid resolution, grid-cell count, training time, inference time per step, peak memory, and rollout error change? Please provide controlled scaling curves and a component-level runtime breakdown on matched hardware, including at least one 3D experiment and PGS.

  3. Can the authors add Sand-3D and Goop-3D experiments using either the full datasets or clearly defined representative subsets? Please include a no-persistence ablation and disclose the exact subset construction, training budget, checkpoint-selection rule, and number of training seeds for every compared method.

8 comments

Thank you for the careful review and helpful suggestions. We particularly appreciate your recognition of the persistent Eulerian state as a meaningful extension of hybrid particle–grid simulators. We also appreciate your positive assessment of our memory ablations and intervention experiments.

Your suggestions on cross-material generalization and 3D scalability have motivated us to substantially expand our experimental evaluation. We have conducted additional experiments to investigate these aspects and present the detailed results below.

Response to W1: Cross-Material Generalization and Scaling

Thank you for raising this important point. To address your concerns, we conducted additional controlled experiments and added them to the camera-ready version, with full details presented below:

  • Cross-material and viscosity generalization in Q1: Joint training and transfer across Water-3D, Sand-3D, and Goop-3D show that embedding-only tuning suffices for intermediate-viscosity interpolation, while few-shot end-to-end tuning adapts well to extrapolated granular dynamics. On the 3D Fluid Cube dataset, continuous viscosity conditioning achieves lower interpolation error than the in-distribution baseline and stable high-viscosity extrapolation.

  • 3D scaling and PGS runtime breakdown in Q2: Scaling experiments on TGV-3D confirm sub-linear growth in inference time and peak memory with particle count, while finer grids consistently improve accuracy. Component profiling on an RTX 4090 shows that particle-side GNNs take over 75% of Full-model runtime, and skipping them during fast steps enables PGS to achieve up to 1.60× speedup in 3D.


Response to W2: 3D Benchmarks and Multi-Seed Evaluation

To address your concerns, we conducted the following experiments.

  • Full Sand-3D and Goop-3D benchmarks and ablations in Q3: CoLEP consistently outperforms CoRGI across all evaluated horizons on both datasets, with particularly substantial long-horizon improvements on Sand-3D. Removing persistent memory degrades accuracy across all evaluated steps on Sand-3D and Goop-3D.

  • 3-seed TGV-3D benchmark in Q3 and 3D PGS in Q2: Multi-seed comparisons on full TGV-3D show that CoLEP outperforms CoRGI and GNS in long-horizon position, kinetic energy, and Sinkhorn metrics with minimal seed variance, while 3D PGS experiments confirm substantial speedups across scales.

Due to tight rebuttal time limits and high 3D training costs, Sand-3D and Goop-3D currently report 1 seed while the remaining seeds are running; we will include full multi-seed ablations in the camera-ready version.

Thank you again for these valuable feedback, which has significantly strengthened our paper.

Responses to Questions

Response to Q1: Material and Viscosity Generalization

Thank you for your insightful question. We fully agree that generalizing to unseen materials and physical parameters is crucial for enhancing the practical utility of neural simulation models. To thoroughly answer Q1, we conducted the following two sets of experiments: (1) cross-material joint training and transfer across Water-3D, Sand-3D, and Goop-3D, and (2) continuous physical parameter interpolation and extrapolation on unseen viscosities.

Part 1: Cross-Material Joint Training and Transfer

To evaluate how CoLEP generalizes and transfers across different material types, we conducted experiments across six settings:

A. Joint Training: Trained jointly on 999 trajectories, with 333 randomly sampled from each of Water-3D, Sand-3D, and Goop-3D.

B. Unseen Sand-3D: Trained jointly on 1,000 trajectories, with 500 randomly sampled each from Water-3D and Goop-3D, and evaluated on unseen Sand-3D.

C. Unseen Goop-3D: Trained jointly on 1,000 trajectories, with 500 randomly sampled each from Water-3D and Sand-3D, and evaluated on unseen Goop-3D.

D. Water-3D Only: Trained exclusively on the full 1,000 training trajectories of Water-3D.

E. Sand-3D Only: Trained exclusively on the full 1,000 training trajectories of Sand-3D.

F. Goop-3D Only: Trained exclusively on the full 1,000 training trajectories of Goop-3D.

Across all six settings from A to F, the models are trained with a comparable total dataset size with about 1,000 trajectories and share the same CoLEP network architecture and training protocol, utilizing a progressive rollout length of 2→42 \rightarrow 4 over a total budget of 20k optimization steps. For evaluation, all models are tested on the official test dataset of the corresponding material, autoregressively predicting 250 rollout steps from the initial 6 frames of each trajectory. Setting D, E, and F serve as single-material baselines to compare with the jointly trained and zero-shot transfer models and measure their effectiveness.

As shown in Table R1-1, despite receiving only one third of the training data per material, Setting A closely matches the single-material baselines on Sand-3D at 0.003920 and Goop-3D at 0.003766, confirming effective shared physical representations. Water-3D incurs a larger error of 0.031059 because rapid fluid splashing accumulates positional drift over 250 rollout steps when per-material data is limited.

For the unseen-material generalization experiments in Settings B and C, we evaluate CoLEP across three adaptation regimes: Zero-Shot, where the pre-trained model is directly tested without parameter updates; Embedding-Only Fine-Tuning, where only the material embedding layer is updated while keeping the backbone frozen; and End-to-End Fine-Tuning, where all parameters are jointly updated. For both fine-tuning regimes, we train on the same 10 target-material trajectories for 300 optimization steps, with results reported in Table R1-2.

As shown in Table R1-2, Goop-3D and Sand-3D behave differently during adaptation. Goop-3D has an intermediate viscosity between water and sand. Because the frozen physics backbone already covers both extremes, tuning only the material embedding is enough to interpolate its behavior, cutting the error from 0.011651 to 0.005592 and nearly matching end-to-end fine-tuning at 0.005548. In contrast, Sand-3D requires extrapolating beyond the viscosities of water and goop. Since the backbone has not seen such granular dynamics, tuning the embedding alone is insufficient, and end-to-end fine-tuning is needed to adapt the physics operators and further reduce the error from 0.009474 to 0.004013.

Test MaterialJoint Model (Setting A)Single-Material Baselines (Settings D / E / F)
Water-3D0.0310590.014467 (Setting D)
Sand-3D0.0039200.002411 (Setting E)
Goop-3D0.0037660.003345 (Setting F)

Table R1-1: Rollout MSE of the jointly trained CoLEP model (Setting A) and single-material CoLEP models (Setting D E and F) across three materials.

Model / Target Unseen MaterialZero-ShotEmbedding-Only Fine-TuningEnd-to-End Fine-Tuning
Setting B (Unseen Sand-3D)0.0143930.0094740.004013
Setting C (Unseen Goop-3D)0.0116510.0055920.005548

Table R1-2: Generalization performance of CoLEP on unseen materials under three adaptation regimes: direct zero-shot inference, fine-tuning only the material embedding layer, and end-to-end fine-tuning.

Part 2: Continuous Viscosity Generalization

To further evaluate CoLEP’s generalization to unseen physical parameters, we conduct viscosity interpolation and extrapolation experiments on the open-source Fluid Cube dataset [1]. We select this benchmark because each of its 100 3D trajectories is governed by a distinct, constant kinematic viscosity ν\nu spanning nearly seven orders of magnitude, providing an ideal testbed for continuous parameter generalization. To systematically construct in-distribution, interpolation, and extrapolation benchmarks without data leakage, we partition the dataset at the trajectory level into five subsets, detailed in Table R1-3: we train on two disjoint viscosity bands, strictly hold out the intermediate gap for interpolation testing, and hold out all extreme high-viscosity trajectories for extrapolation testing.

SplitTrajectoriesKinematic Viscosity (ν\nu) Range & Selection Strategy
Training60Low-viscosity band: ν∈[1.079×10−6, 9.728×10−4]\nu \in [1.079\times 10^{-6},\, 9.728\times 10^{-4}]; High-viscosity band: ν∈[0.005928, 0.896597]\nu \in [0.005928,\, 0.896597]
Validation105 trajectories held out from each training band for checkpoint selection
In-Distribution Test105 additional unseen trajectories held out from each training band
Interpolation Test10Unseen intermediate gap: 0.001≤ν<0.0050.001 \le \nu < 0.005
Extrapolation Test10Unseen high-viscosity trajectories: ν≥1\nu \ge 1

Table R1-3: Trajectory-level dataset splits and kinematic viscosity (ν\nu) distributions for the Fluid Cube generalization benchmark.

To represent unseen continuous parameter, we augment the particle embedding with a continuous physical parameter input. Specifically, we apply a log-transformation followed by Z-score standardization using only the training set statistics: z=log⁡10(ν)−μlog⁡νσlog⁡νz = \frac{\log_{10}(\nu) - \mu_{\log \nu}}{\sigma_{\log \nu}}, where μlog⁡ν\mu_{\log \nu} and σlog⁡ν\sigma_{\log \nu} denote the mean and standard deviation of log⁡10(ν)\log_{10}(\nu) computed across the 60 training trajectories, and concatenate zz directly to the input features of every real particle at each prediction step.

We train CoLEP on the 60 training trajectories and evaluate it across the In-Distribution, Interpolation, and Extrapolation test splits, with the results summarized in Table R1-4.

Test Split100-Step Mean Rollout MSEFrame-100 Position MSE
In-Distribution0.0003150.001900
Interpolation0.0001150.000641
High-Viscosity Extrapolation0.0007500.003044

Table R1-4: 100-step rollout position MSE on the Fluid Cube benchmark across in-distribution, interpolation, and high-viscosity extrapolation regimes.

As shown in Table R1-4, CoLEP demonstrates strong continuous parameter generalization. First, on the in-distribution test set, CoLEP achieves a 100-step mean rollout MSE of 0.0003150.000315, serving as a reference baseline. Second, on the interpolation test set, CoLEP achieves an even lower 100-step mean MSE of 0.0001150.000115, which is 63.5% lower than the in-distribution baseline, confirming smooth and accurate interpolation across the unseen intermediate viscosity gap. Finally, on the high-viscosity extrapolation test set, CoLEP maintains stable, non-divergent rollouts with a 100-step mean MSE of 0.0007500.000750, which is about ∼2.4×\sim 2.4\times the in-distribution baseline, demonstrating meaningful out-of-distribution extrapolation capability.

Reviewer_pX9VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CoLEP adds a persistent latent field on a fixed Eulerian grid to a GNN particle simulator. At each step particle features update the field through a U-Net, the field is read back to refine the particle prediction. Periodic GNN skipping interleaves graph-free steps using adapters on a frozen backbone. On DAM-2D, TGV-2D, Water-3D and a new MovingObstacle-2D, it reports the lowest full-rollout errors against GNS, CoRGI and (in 2D) NeuralMPM.

Critical strengths
  • Persistence was demonstrated to help in a matched ablation.
  • CoLEP leads the published numbers of baselines.
Critical weaknesses
  • Some baselines appear to be under-reproduced, especially DAM-2D and WATER-3D.
  • The function of memory remains underexplored.

See questions for further details.

Questions for authors

Presentation

Clarifications

  1. Eq. 6 advances the field "at every step t>Ht > H", but the first prediction (t=Ht = H) reads out GH+1G^{H+1} (step IV). Is the field also updated at t=Ht = H?

Errata

  1. App. C says the position and kinetic-energy metrics "support different conclusions about the observed trade-off", but Fig. 9 orders the variants identically on both. What does this refer to?

Soundness

Clarifications

  1. How many seeds contributed to Table 1 and Fig. 4a, and what is the spread? Eq. 17 mentions "multiple independently trained seeds" without numbers, and the intervals in Figs. 2 and 6 and Table 3 are across test trajectories, so the ~35% persistence gains on TGV-2D and MovingObstacle-2D are illegible.
  2. What timing protocol underlies Figs. 5a and 8, and were the Fig. 5a baseline latencies measured here? In Fig. 5a, GNS and CoRGI match the CoRGI paper's DAM-2D inference times; on MovingObstacle-2D, which that paper did not run, they take about 45 and 215 ms against CoLEP's 10. GNS and CoRGI are JAX implementations and CoLEP is PyTorch (App. B.5), and two servers are listed (App. B.3).
  3. What training budget, unroll length and BPTT window ("dataset-specific", App. B.5), and what grid resolution and radius (App. B.4), did CoLEP use per dataset, and how long was GNS trained? The GNS reproduction matches the literature at 20 steps, where LagrangeBench's best models reach 3.3×10−53.3\times10^{-5} (GNS-10-128) and 4.4×10−64.4\times10^{-6} (SEGNN-10-64). The CoRGI reproduction does not: its MSE$_{20}$ is 45-56% above CoRGI's published 1.55×10−51.55\times10^{-5} and 3.81×10−63.81\times10^{-6}, and its full-rollout error is 3.2x CoRGI's published DAM-2D value but within 5% on TGV-2D. With 500,000 updates, why the gap on DAM-2D?
  4. Over what horizon is context "accumulated and reused" (Conclusion), and what does the reset-once profile (App. B.7) show?

Errata

  1. NeuralMPM's paper evaluates the same LagrangeBench dam-break data over the full rollout: position MSE 2.082.08-2.91×10−22.91\times10^{-2} and Sinkhorn 2.88×10−42.88\times10^{-4}, the level of CoLEP's 2.57×10−22.57\times10^{-2} and 2.84×10−42.84\times10^{-4}. Table 1's NeuralMPM is 4-6x and 86x worse, from a "validation-selected 10,000-step checkpoint" (App. B.2) where its authors train for 100,000 iterations. Was that checkpoint taken from a full run? And at about 1/16 of CoLEP's latency, does NeuralMPM as published leave any DAM-2D margin or a "favorable accuracy-inference-cost trade-off"?
  2. On Water-3D the GNS paper reports a full-rollout position MSE of 1.01×10−21.01\times10^{-2}. Table 1's GNS (7.88×10−27.88\times10^{-2}) is 7.8x above it in raw units, and CoLEP's 1.44×10−21.44\times10^{-2} is 43% above it. In what units are positions, and does "consistently improves long-horizon rollout accuracy over the evaluated baselines" hold against GNS as published?
  3. Fig. 5b gives a persistence ratio of 0.0304/0.0840=0.3620.0304/0.0840 = 0.362 where Fig. 4a gives 0.305. Its persistent Full bar is close to Table 1 rescaled to fluid nodes, and its transient Full bar to the unrescaled value implied by Fig. 4a. Are the bars on different scopes, from different runs, or from a different checkpoint than Table 1?

Addenda

  1. Both arms of the PGS comparison share one protocol, and the transient model collapses under PGS-1. But its Fast steps run a frozen U-Net through a single-branch adapter trained on App. B.8's short schedule, while NeuralMPM reaches 2.12.1-2.9×10−22.9\times10^{-2} as published. Does a transient model fine-tuned end to end for skipping also collapse?
  2. How much of each margin is persistence? Against CoRGI's published full-rollout errors, CoLEP is at parity on DAM-2D and 30% lower on TGV-2D, while the no-persistence variant is 3x worse and 7% worse, so the TGV-2D lead is attributable to persistence. But on MovingObstacle-2D the ablation beats every baseline, and Water-3D has no ablation.
  3. Is there a test of the Introduction's rationale that retained field information "can help correct deviations in particle predictions", such as recovery after a perturbation? The interventions show only that the state is trajectory-specific and useful; there is no evidence that it corrects.

Informativeness

Clarifications

  1. Does the persistent state help by supplying missing state, by widening the receptive field, or by correcting drift (i.e. the paper's motivation)? What does the evidence actually favor?
  2. What is the utility of a grid memory over, or alongside, a particle-carried latent?
  3. Why adapters on a frozen backbone rather than one model trained with randomly skipped GNN steps?

Addenda

  1. What does GG encode (e.g. smoothed physical fields, the model's own error, non-local context, ...), and how far back does its information reach when probed directly?
  2. Does the persistent state lag at sudden events, e.g. impacts and obstacle reversals in MovingObstacle-2D?
  3. Does the model's reliance on the inherited state change over a rollout?
  4. On Fast steps, what in the inherited state substitutes for the omitted interaction residuals?
  5. What happens when new fluid-obstacle interactions occur during Fast steps?
  6. How does the model's use of memory develop during training, and does it depend on the BPTT window?
14 comments

We thank the reviewer for the thorough review and helpful feedback. Your comments have helped us further strengthen the rigor of our evaluation and clarify the role of memory.

The central contribution of CoLEP goes beyond conventional particle–grid information exchange. By maintaining a persistent Eulerian latent field that co-evolves with moving particles, CoLEP enables spatial information to be accumulated and reused across time steps, improving long-horizon prediction. This persistent field also provides the foundation for Periodic GNN Skipping (PGS), which reduces expensive particle interactions while preserving particle–grid updates.

Following your suggestions, we have revisited our baseline evaluations and conducted additional experiments to better understand the contribution and behavior of persistent memory. We address these points in detail below.

Response to W1: Baseline Reliability and Reproducibility

To address your concerns about baseline reproduction and evaluation reliability, we conducted the following additional experiments and corrections:

  • Baseline reproduction and multi-seed evaluation in Q3, Q5, Q7, and Q8: We extended the main benchmark comparisons to three independent seeds, corrected the NeuralMPM evaluation issue, revisited CoRGI's training budgets and checkpoint selection, and clarified the differences between our GNS baseline and the original implementation.

  • Unified inference benchmarking and evaluation corrections in Q4 and Q9: We re-measured inference times under a unified hardware and timing protocol and identified and corrected inconsistencies in the evaluation settings and reported results.

These revisions strengthen the reliability and reproducibility of our comparisons, while the updated results continue to support CoLEP's predictive advantages.

Response to W2: Understanding Persistent Eulerian Memory

To further investigate the contribution and behavior of persistent memory, we conducted the following experiments and analyses:

  • Extended persistence ablations in Q3 and Q11: We added multi-seed memory ablations and extended the evaluation to Sand-3D and Goop-3D, providing further evidence of the benefits and limitations of Eulerian persistence across different physical systems.

  • Memory intervention and temporal dependence analyses in Q6, Q12, Q13, and Q18: We examined the effects of resetting and freezing the Eulerian state at different rollout stages, providing evidence that persistent memory helps suppress long-term error accumulation and remains useful throughout the trajectory.

  • Memory learning and PGS stability in Q10 and Q21: We investigated how BPTT window length affects memory utilization and evaluated whether end-to-end fine-tuning can improve the skipping stability of a transient-state model.

  • Interpretation of the persistent field in Q14, Q16, and Q19: We further clarified the information carried by the Eulerian state, its advantages over particle-carried memory, and how it supports graph-free prediction steps.

Together, these additional results further support our central contribution. The persistent Eulerian field serves as a continuously evolving spatial memory that improves long-horizon prediction and supports efficient particle simulation through PGS.

Responses to Questions

Response to Q1: Initial Field Update

Yes, the field is also updated at t=Ht=H. The condition “t>Ht\gt H” was a typo and should read “t≥Ht\ge H.” We have corrected this typo in the revised version.


Response to Q2: Appendix C Metric Comparison

This sentence referred to another set of evaluation results where the ranking of the three model variants differed between these two metrics and Sinkhorn distance, which we inadvertently omitted from Appendix C in the original submission. We have added the corresponding figure and analysis to Appendix C in the revised version.

Response to Q3: Multi-Seed Evaluation

In the original submission, Table 1 and Fig. 4a were reported from 1 random seed. To properly quantify training variability, we repeated all experiments in Table 1 across 3 independent random seeds and updated the corresponding results in the Appendix D in the revised version. We also explicitly specified this seed count 33 in Eq. 17 to remove any ambiguity. Table R3-1 below reports the updated 3-seed results. As shown in Table R3-1, CoLEP consistently achieves the lowest mean error across all datasets and metrics, with generally small seed-to-seed variability. These results demonstrate that CoLEP's predictive advantages are consistent across independent training runs.

To verify that the gains from cross-step persistent memory exceed training-seed variability, we also repeated the persistence Eulerian state ablation across 3 random seeds on TGV-2D in Table R3-2. Persistent memory reduces the mean MSEfull\text{MSE}_{\text{full}} on TGV-2D by 25.7%. We will include the complete 3-seed persistence ablation results for TGV-2D, DAM-2D and Moving Obstacle-2D in the updated Figure 4 in the camera-ready version.

DatasetModelMSE₂₀ ↓KE–MSE₂₀ ↓MSE_full ↓KE–MSE_full ↓Sinkhorn_full ↓
DAM-2DGNS(4.47 ± 0.66) × 10⁻⁵(3.29 ± 0.97) × 10⁻⁵(3.17 ± 2.57) × 10⁻¹(1.39 ± 1.29) × 10⁻²(4.34 ± 3.19) × 10⁻²
NeuralMPM(5.27 ± 0.70) × 10⁻⁵(3.53 ± 0.31) × 10⁻⁴(2.82 ± 0.40) × 10⁻²(5.66 ± 1.84) × 10⁻⁴(9.48 ± 6.15) × 10⁻⁴
CoRGI(1.76 ± 0.23) × 10⁻⁵(2.14 ± 1.96) × 10⁻⁵(8.26 ± 5.27) × 10⁻²(2.21 ± 1.99) × 10⁻³(4.83 ± 4.93) × 10⁻³
CoLEP (Ours)(1.31 ± 0.01) × 10⁻⁵(2.66 ± 0.15) × 10⁻⁶(2.16 ± 0.10) × 10⁻²(3.35 ± 0.78) × 10⁻⁵(2.52 ± 0.26) × 10⁻⁴
TGV-2DGNS(7.04 ± 0.28) × 10⁻⁶(2.40 ± 0.26) × 10⁻⁷(5.74 ± 0.04) × 10⁻⁴(2.23 ± 0.23) × 10⁻⁷(9.85 ± 2.56) × 10⁻⁷
NeuralMPM(2.37 ± 0.04) × 10⁻⁵(6.01 ± 2.21) × 10⁻⁶(9.59 ± 0.41) × 10⁻⁴(4.13 ± 1.55) × 10⁻⁶(8.99 ± 1.18) × 10⁻⁶
CoRGI(5.91 ± 0.33) × 10⁻⁶(1.44 ± 0.30) × 10⁻⁷(4.81 ± 0.21) × 10⁻⁴(1.39 ± 0.58) × 10⁻⁷(1.14 ± 1.05) × 10⁻⁶
CoLEP (Ours)(3.55 ± 0.07) × 10⁻⁶(5.96 ± 0.85) × 10⁻⁸(3.23 ± 0.06) × 10⁻⁴(6.09 ± 1.23) × 10⁻⁸(2.04 ± 0.82) × 10⁻⁷
Moving Obstacle-2DGNS(9.78 ± 0.62) × 10⁻⁶(7.24 ± 0.89) × 10⁻⁷(2.03 ± 1.15) × 10⁻²(2.84 ± 0.79) × 10⁻⁵(2.26 ± 1.70) × 10⁻²
NeuralMPM(2.84 ± 0.21) × 10⁻⁵(1.38 ± 0.35) × 10⁻⁶(1.70 ± 0.10) × 10⁻³(2.00 ± 0.24) × 10⁻⁶(9.03 ± 0.46) × 10⁻⁴
CoRGI(6.68 ± 0.19) × 10⁻⁶(4.07 ± 1.02) × 10⁻⁷(3.39 ± 0.26) × 10⁻³(4.23 ± 0.35) × 10⁻⁶(2.39 ± 0.51) × 10⁻³
CoLEP (Ours)(1.61 ± 0.06) × 10⁻⁶(1.52 ± 0.01) × 10⁻⁷(8.92 ± 0.96) × 10⁻⁴(5.58 ± 1.13) × 10⁻⁷(2.47 ± 0.08) × 10⁻⁴
Water-3DGNS(5.63 ± 0.41) × 10⁻⁵(4.97 ± 2.30) × 10⁻⁴(8.75 ± 2.29) × 10⁻²(9.66 ± 6.52) × 10⁻³(1.10 ± 0.75) × 10⁻¹
CoRGI(4.00 ± 0.30) × 10⁻⁵(3.28 ± 0.13) × 10⁻⁴(6.79 ± 1.92) × 10⁻²(3.27 ± 1.88) × 10⁻³(6.53 ± 5.56) × 10⁻²
CoLEP (Ours)(3.72 ± 0.15) × 10⁻⁶(1.91 ± 0.68) × 10⁻⁵(1.42 ± 0.02) × 10⁻²(3.87 ± 0.69) × 10⁻⁵(6.28 ± 2.65) × 10⁻³

Table R3-1: Quantitative rollout accuracy across four benchmarks evaluated over 3 independent random seeds (mean ± SD).

In Table 1 of our original submission, the 20-step errors on MovingObstacle-2D of NeuralMPM were computed only on the first 20-step window of each test trajectory. In the updated Table R3-1, we fixed this protocol to average the errors over all 20-step windows across the full trajectory. Because later windows include strong collisions with the moving obstacle, the new 20-step errors are much larger. We have also updated the related discussion in the revised paper.

SeedCoLEP without Persistent MemoryFull CoLEP
14.87 × 10⁻⁴3.17 × 10⁻⁴
24.16 × 10⁻⁴3.24 × 10⁻⁴
34.02 × 10⁻⁴3.28 × 10⁻⁴
Mean ± SD(4.35 ± 0.45) × 10⁻⁴(3.23 ± 0.06) × 10⁻⁴

Table R3-2: Per-seed MSEfull\text{MSE}_{\text{full}} comparison for the Eulerian state persistence ablation on TGV-2D


Response to Q4: Unified Inference Timing

In the original figure 5a, the DAM-2D inference times for GNS at 11.500 ms per step and CoRGI at 11.300 ms per step were taken from the CoRGI paper, as our measurements were slower at 18.693 ms and 42.787 ms per step and we aimed to avoid underestimating the baselines. All the following results across both datasets were measured on a single A100 GPU with a batch size of 1, while the 4090 server listed in Appendix B.3 was used exclusively for model training.

The high latencies of roughly 214.3 ms per step for CoRGI and 44.9 ms per step for GNS on MovingObstacle-2D resulted from recompilation overhead triggered by neighbor list overflow. Specifically, the original implementations initialized neighbor list capacities using training statistics, but accumulated prediction errors during long rollouts caused local particle clustering that exceeded this preset limit. Under default settings, each overflow forced dynamic list expansion and JAX recompilation, which were included in the recorded runtime.

To ensure a fair comparison, we re-evaluated all methods under a unified protocol on the same A100 GPU with a batch size of 1 and TF32 precision enabled across all frameworks. For each dataset, we fixed 50 randomly sampled R20 windows, warmed them up completely, and measured the average per-step wall-clock time over a single pass, covering neighbor updates, feature construction, inference, boundary handling, integration, and device synchronization while excluding data loading. For GNS and CoRGI, we profiled the test trajectories beforehand to preallocate a fixed neighbor list capacity covering peak neighbor counts, ensuring zero expansions or recompilations during timing.

The updated latencies in milliseconds per step under this unified protocol are shown in Table R4-1 below:

MethodDAM-2DMovingObstacle-2D
GNS13.11337.526
CoRGI28.46369.075
NeuralMPM0.8301.477
CoLEP Full (Ours)8.78611.386
PGS-1 (Ours)6.7617.648
PGS-2 (Ours)6.1296.239

Table R4-1: Unified per-step inference latency on DAM-2D and MovingObstacle-2D (ms/step, NVIDIA A100).

Without recompilation overhead, the latency of CoRGI on MovingObstacle-2D drops to 69.075 ms per step, confirming that our methods remain substantially faster than GNS and CoRGI on both datasets. We will update Figures 5a and 8 with these unified results and document the server roles and timing protocol in Appendix B.5.


Response to Q5: Training Settings and CoRGI Reproduction

First, we show the CoLEP training settings in Table R5-1, which lists the total optimizer updates and rollout curriculum per dataset. Here, RR denotes the training unroll length, with full BPTT applied across all RR steps. Within each transition window, the probability of sampling the longer rollout increases linearly, and all ranges are inclusive.

Second, we show the CoLEP grid and radius settings in Table R5-2, which lists the Eulerian grid resolutions and particle interaction radii for each dataset.

For GNS training, we used the official LagrangeBench implementation, trained each seed for 500,000 optimizer updates per dataset, and evaluated the checkpoint with the lowest validation 20-step position MSE.

Finally, regarding the CoRGI reproduction gap on DAM-2D, while the original paper reports all other training settings, it does not specify the total number of optimizer updates. Using their official implementation and validation MSE20\text{MSE}_{20} checkpoint selection rule, we therefore tested two training budgets of 150,000 and 500,000 updates. 500,000 is inferred from their official codebase configurations and 150,000 is estimated from their reported training runtime. At 150,000 updates, one seed reproduced the published accuracy well, reaching a full-rollout position MSE of 0.03445280.0344528 at step 115k compared to the reported 2.84×10−22.84 \times 10^{-2}.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.