PAINET: A Principled Efficient Transformer for 3D Dynamics Modeling

Junheng Tao; Kai Yang; Qitian Wu; Wanyu Wang; Yuqi Huang

arxiv: 2510.04233 · v2 · submitted 2025-10-05 · 💻 cs.LG · cs.AI

PAINET: A Principled Efficient Transformer for 3D Dynamics Modeling

Kai Yang , Yuqi Huang , Junheng Tao , Wanyu Wang , Qitian Wu This is my paper

Pith reviewed 2026-05-18 09:54 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords 3D dynamics modelingtransformermulti-body systemsenergy minimizationattention networktrajectory predictionequivariant model

0 comments

The pith

PAINET derives transformer attention from energy minimization trajectories to model unobserved interactions in 3D multi-body dynamics.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to predict how groups of objects move together in three dimensions when some of their influences on each other are not directly recorded. It builds a transformer whose attention step follows the path a physical system would take while lowering its total energy. This choice is paired with a decoder that keeps predictions unchanged under rotations and translations of the whole scene. If the approach holds, simulators could produce accurate trajectories for molecules, proteins, or human bodies without first listing every possible force or contact. Experiments on motion-capture, molecular-dynamics, and protein datasets report lower errors than earlier methods at similar running cost.

Core claim

PAINET is a transformer for 3D dynamics that contains a physics-inspired attention network derived from the minimization trajectory of an energy function together with a parallel decoder that keeps the overall mapping unchanged under rotations and translations. The design lets the model learn all-pair interactions among bodies even when those interactions are not explicitly observed, and it produces lower prediction errors than prior models on human motion capture, molecular dynamics, and large-scale protein simulation benchmarks.

What carries the argument

The physics-inspired attention network that obtains its weights from the trajectory of minimizing an energy function, together with the parallel decoder that keeps rotational and translational symmetry.

If this is right

The model can still forecast trajectories when some interactions between bodies remain unobserved in the input data.
Equivariance under rotations and translations is maintained while inference runs in parallel.
Error reductions between 4.7 percent and 41.5 percent appear across motion-capture, molecular, and protein benchmarks at comparable memory and time cost.
The same architecture applies without change to systems of different sizes and different physical domains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Energy-derived attention may transfer to other symmetry-constrained simulation tasks outside the three domains tested.
The parallel-decoder choice could reduce memory growth when the number of bodies increases.
Replacing the unspecified energy function with a domain-specific one might further tighten long-horizon forecasts.

Load-bearing premise

That attention weights taken from the path of energy minimization will still respect geometric symmetries and will correctly handle interaction types never shown during training.

What would settle it

Measuring whether prediction error on a new multi-body scene whose interaction pattern differs from all training examples becomes no better than that of a standard transformer without the energy-minimization step.

Figures

Figures reproduced from arXiv: 2510.04233 by Junheng Tao, Kai Yang, Qitian Wu, Wanyu Wang, Yuqi Huang.

**Figure 2.** Figure 2: Representative snapshots of aspirin molecular dynamics: the top row shows the ground [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Figure 3: Illustration for three types of equivariance, including rotation, translation and permutation. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

**Figure 4.** Figure 4: Ablation studies w.r.t. learnable pairwise mappings in the attention network and the [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗

**Figure 5.** Figure 5: Ablation studies w.r.t. the number of decoding layers on Motion Capture. [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗

**Figure 6.** Figure 6: Ablation studies w.r.t. the number of attention layers on Molecular Dynamics. [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗

**Figure 7.** Figure 7: Scalability test including inference time and GPU memory cost w.r.t. time steps and [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗

**Figure 8.** Figure 8: Representative snapshots of toluene molecular dynamics, initialized at snapshot 60666.0 [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗

**Figure 9.** Figure 9: Representative snapshots of toluene molecular dynamics starting at 79100.0 ps. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

**Figure 10.** Figure 10: Representative snapshots of salicylic molecular dynamics starting at 56292.0 ps. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗

**Figure 11.** Figure 11: Representative snapshots of salicylic molecular dynamics starting at 77169.0 ps. [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗

read the original abstract

Modeling 3D dynamics is a fundamental problem in multi-body systems across scientific and engineering domains and has important practical implications in object trajectory prediction and simulation. While recent GNN-based approaches have achieved strong performance by enforcing geometric symmetries, encoding high-order features or incorporating neural-ODE mechanics, they typically depend on explicitly observed structures and inherently fail to capture the unobserved interactions that are crucial to complex physical behaviors and dynamics mechanism. In this paper, we propose PAINET, a principled SE(3)-equivariant transformer for learning all-pair interactions in multi-body systems. The model comprises: (1) a novel physics-inspired attention network derived from the minimization trajectory of an energy function, and (2) a parallel decoder that preserves equivariance while enabling efficient inference. Empirical results on diverse real-world benchmarks, including human motion capture, molecular dynamics, and large-scale protein simulations, show that PAINET consistently outperforms recently proposed models, yielding 4.7% to 41.5% error reductions in 3D dynamics prediction with comparable computation costs in terms of time and memory. Our codes, baseline models and datasets are available at https://github.com/Icarus1411/PAINET.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PAINET tries an energy-minimization attention inside an SE(3) transformer but the derivation is not shown explicitly enough to back the principled claim.

read the letter

Hi colleague, The main thing here is that PAINET builds an SE(3)-equivariant transformer for 3D multi-body dynamics with attention weights taken from the trajectory of minimizing an energy function, plus a parallel decoder meant to keep equivariance while speeding up inference. It targets cases where interaction graphs are not observed upfront. The experiments are the strongest part. Tests on human motion capture, molecular dynamics, and large protein simulations show error reductions from 4.7% to 41.5% against recent GNN and neural-ODE baselines, and the authors release code and data. That range of domains and the public artifacts make the results easier to check. The parallel decoder is a reasonable engineering choice for efficiency. What looks new is the specific combination of the energy-trajectory attention with the equivariant transformer and decoder. The soft spot is the core modeling step. The abstract and description say the attention comes from energy minimization, yet there is no closed-form energy function, no explicit minimization steps, and no proof that the resulting weights stay SE(3)-equivariant on unseen interactions. Without that, it is difficult to tell how much is principled derivation versus tuned attention. The write-up also skips ablations on the energy component and any statistical tests, so the gains could be sensitive to implementation details. This paper is for researchers who work on simulation in robotics, molecular modeling, or structural biology and want architectures that learn hidden pairwise interactions without graphs. Someone already following equivariant transformers or physics-informed layers would find the benchmarks worth reading. It deserves peer review. The empirical coverage and the attempt at a new attention mechanism give it enough substance for referees to evaluate, even if the derivation section needs tightening.

Referee Report

2 major / 2 minor

Summary. The paper introduces PAINET, a SE(3)-equivariant transformer for modeling 3D dynamics in multi-body systems. It consists of a physics-inspired attention network claimed to be derived from the minimization trajectory of an energy function, together with a parallel decoder that maintains equivariance for efficient inference. The model is evaluated on human motion capture, molecular dynamics, and large-scale protein simulation benchmarks, reporting error reductions of 4.7% to 41.5% relative to recent baselines while maintaining comparable computational cost.

Significance. If the central derivation can be made explicit and the reported gains prove robust under controlled training protocols and statistical testing, PAINET would offer a useful advance in equivariant architectures for physical dynamics by linking attention to an energy-minimization principle. The public release of code, baselines, and datasets is a clear strength that supports reproducibility.

major comments (2)

[§3] §3 (Physics-Inspired Attention Network): The central claim that the attention mechanism is 'derived from the minimization trajectory of an energy function' is load-bearing for the 'principled' distinction from standard equivariant transformers, yet the manuscript provides neither the explicit form of the energy function E nor the closed-form steps (gradient flow, variational derivation, or iterative minimization) that produce the attention weights. Without this, it is impossible to confirm that the resulting attention is guaranteed to be SE(3)-equivariant for unobserved interactions rather than empirically fitted.
[§4] §4 (Experiments): The reported error reductions (4.7%–41.5%) are presented without details on hyperparameter search, number of random seeds, statistical significance tests, or an ablation isolating the energy-function component. These omissions make it difficult to assess whether the gains are robust or sensitive to implementation choices.

minor comments (2)

[Abstract] The abstract and introduction would benefit from a concise statement of the exact energy function and the derivation steps, even if the full algebra appears later.
[§3.3] Notation for the parallel decoder and its equivariance preservation could be clarified with a short diagram or pseudocode to aid readers unfamiliar with the architecture.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We sincerely thank the referee for the detailed and constructive feedback on our manuscript. The comments highlight important areas for improvement in clarity and experimental rigor. We address each major comment below and commit to making the necessary revisions to enhance the paper.

read point-by-point responses

Referee: [§3] §3 (Physics-Inspired Attention Network): The central claim that the attention mechanism is 'derived from the minimization trajectory of an energy function' is load-bearing for the 'principled' distinction from standard equivariant transformers, yet the manuscript provides neither the explicit form of the energy function E nor the closed-form steps (gradient flow, variational derivation, or iterative minimization) that produce the attention weights. Without this, it is impossible to confirm that the resulting attention is guaranteed to be SE(3)-equivariant for unobserved interactions rather than empirically fitted.

Authors: We appreciate the referee's emphasis on this critical aspect. The manuscript describes the attention as derived from energy minimization, but we agree the explicit form and steps should be elaborated for clarity. In the revised manuscript, we will add the explicit energy function E, which is constructed as a sum of pairwise interaction potentials that are SE(3)-invariant, and detail the iterative minimization process (specifically, a single gradient descent step on the energy with respect to the attention logits) that yields the attention weights. This ensures the equivariance holds for unobserved interactions by construction, as the energy depends only on relative coordinates. We will include a proof that the attention mechanism inherits the SE(3)-equivariance from the energy function. revision: yes
Referee: [§4] §4 (Experiments): The reported error reductions (4.7%–41.5%) are presented without details on hyperparameter search, number of random seeds, statistical significance tests, or an ablation isolating the energy-function component. These omissions make it difficult to assess whether the gains are robust or sensitive to implementation choices.

Authors: We agree that additional experimental details are necessary for reproducibility and to substantiate the claims. We will report the hyperparameter search procedure (grid search over learning rates, batch sizes, and dimensions based on validation), the use of 5 random seeds with mean and standard deviation, and statistical significance via paired t-tests (p < 0.01). We will also add an ablation study isolating the energy-minimization component by comparing against a standard attention variant. These updates, including revised tables, will be incorporated into §4 and the supplement. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The provided abstract and context describe PAINET's attention network as 'derived from the minimization trajectory of an energy function' without exhibiting any explicit equations, parameter fitting steps, or self-citations that reduce the claimed derivation to the model's inputs or training data by construction. No load-bearing self-citation, uniqueness theorem, or ansatz smuggling is quoted. The empirical benchmark improvements are presented as validation rather than the source of the architecture itself. The derivation chain remains self-contained against external benchmarks and does not collapse into fitted inputs renamed as predictions or self-definitional loops.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The model rests on the assumption that pairwise interactions dominate and can be recovered via an energy-minimization process whose functional form is not fully specified in the abstract; no new particles or forces are postulated, but the energy function itself functions as an implicit modeling choice whose parameters may be learned.

free parameters (1)

energy_function_parameters
The precise parameterization of the energy function whose minimization trajectory defines the attention is not detailed; any coefficients or functional forms chosen to produce the attention weights count as free parameters fitted during training.

axioms (1)

domain assumption SE(3) equivariance must be preserved by both the attention and the decoder
Invoked when stating that the parallel decoder preserves equivariance; this is a standard symmetry requirement for 3D physical systems but is treated as given rather than re-derived.

pith-pipeline@v0.9.0 · 5746 in / 1587 out tokens · 26049 ms · 2026-05-18T09:54:01.497197+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

Theorem 1 ... iterative updating rule ... yields a descent step on the energy ... ω_ij = ∂ρ_ij(h²)/∂h² ... f_ij(h²)=a_ij-2b_ij h²
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

ρ_ij(h²)=a_ij h² - b_ij h⁴ ... Landau-Ginzburg potential energy form

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 1 internal anchor

[1]

Nicol`o Defenu, Tobias Donner, Tommaso Macr`ı, Guido Pagano, Stefano Ruffo, and Andrea Trom- bettoni

Accessed: 2025-05-09. Nicol`o Defenu, Tobias Donner, Tommaso Macr`ı, Guido Pagano, Stefano Ruffo, and Andrea Trom- bettoni. Long-range interacting quantum systems.Reviews of Modern Physics, 95(3):035002,

work page 2025
[2]

Long range interac- tions in nanoscale science.Reviews of Modern Physics, 82(2):1887–1944,

Roger H French, V Adrian Parsegian, Rudolf Podgornik, Rick F Rajter, Anand Jagota, Jian Luo, Dilip Asthagiri, Manoj K Chaudhury, Yet-ming Chiang, Steve Granick, et al. Long range interac- tions in nanoscale science.Reviews of Modern Physics, 82(2):1887–1944,

work page 1944
[3]

Se (3)-transformers: 3d roto- translation equivariant attention networks.Advances in neural information processing systems, 33:1970–1981,

Fabian Fuchs, Daniel Worrall, V olker Fischer, and Max Welling. Se (3)-transformers: 3d roto- translation equivariant attention networks.Advances in neural information processing systems, 33:1970–1981,

work page 1970
[4]

A survey of geometric graph neural networks: Data structures, models and applications.arXiv preprint arXiv:2403.00485, 2024a

Jiaqi Han, Jiacheng Cen, Liming Wu, Zongzhao Li, Xiangzhe Kong, Rui Jiao, Ziyang Yu, Tingyang Xu, Fandi Wu, Zihe Wang, et al. A survey of geometric graph neural networks: Data structures, models and applications.arXiv preprint arXiv:2403.00485, 2024a. Jiaqi Han, Minkai Xu, Aaron Lou, Haotian Ye, and Stefano Ermon. Geometric trajectory diffusion models, 20...

work page arXiv
[5]

Equivariant graph mechanics networks with constraints.arXiv preprint arXiv:2203.06442,

Wenbing Huang, Jiaqi Han, Yu Rong, Tingyang Xu, Fuchun Sun, and Junzhou Huang. Equivariant graph mechanics networks with constraints.arXiv preprint arXiv:2203.06442,

work page arXiv
[6]

George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

URLhttps://arxiv.org/abs/2410.06366. George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440,

work page arXiv
[7]

Semi-Supervised Classification with Graph Convolutional Networks

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net- works.arXiv preprint arXiv:1609.02907,

work page internal anchor Pith review Pith/arXiv arXiv
[8]

Equivariant flows: sampling configurations for multi- body systems with symmetric energies.arXiv preprint arXiv:1910.00753,

Jonas K ¨ohler, Leon Klein, and Frank No ´e. Equivariant flows: sampling configurations for multi- body systems with symmetric energies.arXiv preprint arXiv:1910.00753,

work page arXiv 1910
[9]

Learning to simulate complex physics with graph networks.arXiv preprint arXiv:2002.09405,

Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Pe- ter W Battaglia. Learning to simulate complex physics with graph networks.arXiv preprint arXiv:2002.09405,

work page arXiv 2002
[10]

Molecular dynamics trajectory for benchmarking mdanalysis

Sean Seyler and Oliver Beckstein. Molecular dynamics trajectory for benchmarking mdanalysis. URL: https://figshare. com/articles/Molecular dynamics trajectory for benchmarking MDAnaly- sis/5108170, doi, 10(m9):7,

work page arXiv
[11]

Aaron Taudt, Axel Arnold, and J ¨urgen Pleiss

URL https://arxiv.org/abs/2411.01600. Aaron Taudt, Axel Arnold, and J ¨urgen Pleiss. Simulation of protein association: Kinetic pathways towards crystal contacts.Physical Review E, 91(3):033311,

work page arXiv
[12]

Thiemann, Thiago Resch¨ utzegger, Massimiliano Esposito, Tseden Taddese, Juan D

URLhttps://arxiv.org/abs/2503.23794. Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds.arXiv preprint arXiv:1802.08219,

work page arXiv
[13]

arXiv preprint arXiv:2407.03925 , year=

URL https://arxiv.org/abs/2407.03925. Hao Wu, Fan Xu, Yifan Duan, Ziwei Niu, Weiyan Wang, Gaofeng Lu, Kun Wang, Yuxuan Liang, and Yang Wang. Spatio-temporal fluid dynamics modeling via physical-awareness and parameter diffusion guidance,

work page arXiv
[14]

12 Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang

URLhttps://arxiv.org/abs/2403.13850. 12 Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geo- metric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923,

work page arXiv
[15]

Equivariant graph neural operator for modeling 3d dynamics.arXiv preprint arXiv:2401.11037,

Minkai Xu, Jiaqi Han, Aaron Lou, Jean Kossaifi, Arvind Ramanathan, Kamyar Azizzadenesheli, Jure Leskovec, Stefano Ermon, and Anima Anandkumar. Equivariant graph neural operator for modeling 3d dynamics.arXiv preprint arXiv:2401.11037,

work page arXiv
[16]

3D Dynamics Prediction.Predicting 3D dynamics—such as particle motion, robotic trajectories, or molecular interactions—is a fundational challenge in physics simulation and robotics

A RELATEDWORKS We briefly discuss the relevant literature to include more background. 3D Dynamics Prediction.Predicting 3D dynamics—such as particle motion, robotic trajectories, or molecular interactions—is a fundational challenge in physics simulation and robotics. Traditional physics-based models offer interpretability but often fall short when modelin...

work page 2021
[17]

Some subsequent works focus on the efficiency of EGNNs through architectural optimizations (Zhang et al., 2024)

further improve physical fidelity by preserving Euclidean symmetries, enabling more accurate and generalizable 3D dynamics predictions. Some subsequent works focus on the efficiency of EGNNs through architectural optimizations (Zhang et al., 2024). Message Passing Neural Networks.Graph neural networks (GNNs) (Gilmer et al., 2017; Kipf & Welling,

work page 2024
[18]

provide a general framework for learning over relational structures and have rec- ognized as a powerful method for simulating physical systems. One line of related works har- ness physical priors to design more expressive message-passing operators to encode system interac- tions (Mrowca et al., 2018; Shi et al., 2024; Viswanath et al.,

work page 2018
[19]

or incorporate classical physical mechanics into the architecture (Sanchez-Gonzalez et al., 2019). A parallel line of works focus on encoding the symmetries of Euclidean space, i.e., equivariance, as an inductive bias into the architec- tures, including translation equivariance (Ummenhofer et al., 2019; Sanchez-Gonzalez et al., 2020; Pfaff et al., 2021), ...

work page 2019
[20]

Further refinements exploit local coordinate frames to process higher- order geometric features (Liu et al., 2022; Du et al., 2022; Han et al., 2024b; Cen et al.,

or equivariant message passing (Satorras et al., 2021; Huang et al., 2022; Thiemann et al., 2025). Further refinements exploit local coordinate frames to process higher- order geometric features (Liu et al., 2022; Du et al., 2022; Han et al., 2024b; Cen et al.,

work page 2021
[21]

becomes a function oftand its optimum is achieved if and only ifH (t) =H. Such a fact yields that E(H(t), t;{ρ ij})≥E(H (t), t+ 1;{ρ ij}).(19) The result of the main theorem follows by noting thatE(H (t), t;{ρ ij})≥E(H (t+1), t;{ρ ij})≥ E(H(t+1), t+ 1;{ρ ij}). D PROOF FOREQUIVARIANCE OVERPAINET We provide a formal proof that our model architecture preserv...

work page 2003
[22]

Following prior work, joints are modeled as graph nodes, with connections representing physical or kinematic constraints

Each snapshot in every trajectory consists of 3D coordinates of 31 human joints. Following prior work, joints are modeled as graph nodes, with connections representing physical or kinematic constraints. The model is conditioned on the initial positions and velocities of joints, and trained to predict future joint positions over a 100-step horizon. Trainin...

work page 2017
[23]

Edge features include a combination of atomic types, bond types, and hop distance between connected atoms

and aug- ment the connectivity by adding 2-hop neighbors. Edge features include a combination of atomic types, bond types, and hop distance between connected atoms. For both training and evaluation, we setT= 8inS2Ttasks. Namely, for each initial position, the model needs to predict the positions in the next 8 time step. E.4 PROTEINDYNAMICS(ADK) We evaluat...

work page 2017

[1] [1]

Nicol`o Defenu, Tobias Donner, Tommaso Macr`ı, Guido Pagano, Stefano Ruffo, and Andrea Trom- bettoni

Accessed: 2025-05-09. Nicol`o Defenu, Tobias Donner, Tommaso Macr`ı, Guido Pagano, Stefano Ruffo, and Andrea Trom- bettoni. Long-range interacting quantum systems.Reviews of Modern Physics, 95(3):035002,

work page 2025

[2] [2]

Long range interac- tions in nanoscale science.Reviews of Modern Physics, 82(2):1887–1944,

Roger H French, V Adrian Parsegian, Rudolf Podgornik, Rick F Rajter, Anand Jagota, Jian Luo, Dilip Asthagiri, Manoj K Chaudhury, Yet-ming Chiang, Steve Granick, et al. Long range interac- tions in nanoscale science.Reviews of Modern Physics, 82(2):1887–1944,

work page 1944

[3] [3]

Se (3)-transformers: 3d roto- translation equivariant attention networks.Advances in neural information processing systems, 33:1970–1981,

Fabian Fuchs, Daniel Worrall, V olker Fischer, and Max Welling. Se (3)-transformers: 3d roto- translation equivariant attention networks.Advances in neural information processing systems, 33:1970–1981,

work page 1970

[4] [4]

A survey of geometric graph neural networks: Data structures, models and applications.arXiv preprint arXiv:2403.00485, 2024a

Jiaqi Han, Jiacheng Cen, Liming Wu, Zongzhao Li, Xiangzhe Kong, Rui Jiao, Ziyang Yu, Tingyang Xu, Fandi Wu, Zihe Wang, et al. A survey of geometric graph neural networks: Data structures, models and applications.arXiv preprint arXiv:2403.00485, 2024a. Jiaqi Han, Minkai Xu, Aaron Lou, Haotian Ye, and Stefano Ermon. Geometric trajectory diffusion models, 20...

work page arXiv

[5] [5]

Equivariant graph mechanics networks with constraints.arXiv preprint arXiv:2203.06442,

Wenbing Huang, Jiaqi Han, Yu Rong, Tingyang Xu, Fuchun Sun, and Junzhou Huang. Equivariant graph mechanics networks with constraints.arXiv preprint arXiv:2203.06442,

work page arXiv

[6] [6]

George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

URLhttps://arxiv.org/abs/2410.06366. George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440,

work page arXiv

[7] [7]

Semi-Supervised Classification with Graph Convolutional Networks

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net- works.arXiv preprint arXiv:1609.02907,

work page internal anchor Pith review Pith/arXiv arXiv

[8] [8]

Equivariant flows: sampling configurations for multi- body systems with symmetric energies.arXiv preprint arXiv:1910.00753,

Jonas K ¨ohler, Leon Klein, and Frank No ´e. Equivariant flows: sampling configurations for multi- body systems with symmetric energies.arXiv preprint arXiv:1910.00753,

work page arXiv 1910

[9] [9]

Learning to simulate complex physics with graph networks.arXiv preprint arXiv:2002.09405,

Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Pe- ter W Battaglia. Learning to simulate complex physics with graph networks.arXiv preprint arXiv:2002.09405,

work page arXiv 2002

[10] [10]

Molecular dynamics trajectory for benchmarking mdanalysis

Sean Seyler and Oliver Beckstein. Molecular dynamics trajectory for benchmarking mdanalysis. URL: https://figshare. com/articles/Molecular dynamics trajectory for benchmarking MDAnaly- sis/5108170, doi, 10(m9):7,

work page arXiv

[11] [11]

Aaron Taudt, Axel Arnold, and J ¨urgen Pleiss

URL https://arxiv.org/abs/2411.01600. Aaron Taudt, Axel Arnold, and J ¨urgen Pleiss. Simulation of protein association: Kinetic pathways towards crystal contacts.Physical Review E, 91(3):033311,

work page arXiv

[12] [12]

Thiemann, Thiago Resch¨ utzegger, Massimiliano Esposito, Tseden Taddese, Juan D

URLhttps://arxiv.org/abs/2503.23794. Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds.arXiv preprint arXiv:1802.08219,

work page arXiv

[13] [13]

arXiv preprint arXiv:2407.03925 , year=

URL https://arxiv.org/abs/2407.03925. Hao Wu, Fan Xu, Yifan Duan, Ziwei Niu, Weiyan Wang, Gaofeng Lu, Kun Wang, Yuxuan Liang, and Yang Wang. Spatio-temporal fluid dynamics modeling via physical-awareness and parameter diffusion guidance,

work page arXiv

[14] [14]

12 Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang

URLhttps://arxiv.org/abs/2403.13850. 12 Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geo- metric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923,

work page arXiv

[15] [15]

Equivariant graph neural operator for modeling 3d dynamics.arXiv preprint arXiv:2401.11037,

Minkai Xu, Jiaqi Han, Aaron Lou, Jean Kossaifi, Arvind Ramanathan, Kamyar Azizzadenesheli, Jure Leskovec, Stefano Ermon, and Anima Anandkumar. Equivariant graph neural operator for modeling 3d dynamics.arXiv preprint arXiv:2401.11037,

work page arXiv

[16] [16]

3D Dynamics Prediction.Predicting 3D dynamics—such as particle motion, robotic trajectories, or molecular interactions—is a fundational challenge in physics simulation and robotics

A RELATEDWORKS We briefly discuss the relevant literature to include more background. 3D Dynamics Prediction.Predicting 3D dynamics—such as particle motion, robotic trajectories, or molecular interactions—is a fundational challenge in physics simulation and robotics. Traditional physics-based models offer interpretability but often fall short when modelin...

work page 2021

[17] [17]

Some subsequent works focus on the efficiency of EGNNs through architectural optimizations (Zhang et al., 2024)

further improve physical fidelity by preserving Euclidean symmetries, enabling more accurate and generalizable 3D dynamics predictions. Some subsequent works focus on the efficiency of EGNNs through architectural optimizations (Zhang et al., 2024). Message Passing Neural Networks.Graph neural networks (GNNs) (Gilmer et al., 2017; Kipf & Welling,

work page 2024

[18] [18]

provide a general framework for learning over relational structures and have rec- ognized as a powerful method for simulating physical systems. One line of related works har- ness physical priors to design more expressive message-passing operators to encode system interac- tions (Mrowca et al., 2018; Shi et al., 2024; Viswanath et al.,

work page 2018

[19] [19]

or incorporate classical physical mechanics into the architecture (Sanchez-Gonzalez et al., 2019). A parallel line of works focus on encoding the symmetries of Euclidean space, i.e., equivariance, as an inductive bias into the architec- tures, including translation equivariance (Ummenhofer et al., 2019; Sanchez-Gonzalez et al., 2020; Pfaff et al., 2021), ...

work page 2019

[20] [20]

Further refinements exploit local coordinate frames to process higher- order geometric features (Liu et al., 2022; Du et al., 2022; Han et al., 2024b; Cen et al.,

or equivariant message passing (Satorras et al., 2021; Huang et al., 2022; Thiemann et al., 2025). Further refinements exploit local coordinate frames to process higher- order geometric features (Liu et al., 2022; Du et al., 2022; Han et al., 2024b; Cen et al.,

work page 2021

[21] [21]

becomes a function oftand its optimum is achieved if and only ifH (t) =H. Such a fact yields that E(H(t), t;{ρ ij})≥E(H (t), t+ 1;{ρ ij}).(19) The result of the main theorem follows by noting thatE(H (t), t;{ρ ij})≥E(H (t+1), t;{ρ ij})≥ E(H(t+1), t+ 1;{ρ ij}). D PROOF FOREQUIVARIANCE OVERPAINET We provide a formal proof that our model architecture preserv...

work page 2003

[22] [22]

Following prior work, joints are modeled as graph nodes, with connections representing physical or kinematic constraints

Each snapshot in every trajectory consists of 3D coordinates of 31 human joints. Following prior work, joints are modeled as graph nodes, with connections representing physical or kinematic constraints. The model is conditioned on the initial positions and velocities of joints, and trained to predict future joint positions over a 100-step horizon. Trainin...

work page 2017

[23] [23]

Edge features include a combination of atomic types, bond types, and hop distance between connected atoms

and aug- ment the connectivity by adding 2-hop neighbors. Edge features include a combination of atomic types, bond types, and hop distance between connected atoms. For both training and evaluation, we setT= 8inS2Ttasks. Namely, for each initial position, the model needs to predict the positions in the next 8 time step. E.4 PROTEINDYNAMICS(ADK) We evaluat...

work page 2017