Integrating Physics and Topology in Neural Networks for Learning Rigid Body Dynamics

Amaury Wei; Olga Fink

REVIEW 3 major objections 6 minor 1 cited by

Integrating Physics and Topology in Neural Networks for Learning Rigid Body Dynamics

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A topological network with physics-guided messages learns rigid-body collisions more accurately than graph simulators.

desk verdict A genuinely new architecture for learned rigid-body dynamics with clean ablations; the SOTA claim is plausible but hinges on a FIGNet comparison that needs supplement verification. read the letter →

arxiv 2411.11467 v3 pith:KCKQRRYG submitted 2024-11-18 cs.LG

Amaury Wei , Olga Fink This is my paper · ORCID

classification cs.LG

keywords rigidbodydynamicstopologicaldeeplearningcombinatorialcomplexphysics-informedmessagepassinggraphneuralnetworkscollisionsimulationout-of-distributiongeneralizationautoregressiverollout

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Graph-based simulators represent a rigid-body scene as nodes and edges, which discards the surfaces, the objects, and the contact geometry that actually determine collisions. The paper proposes HOPNet, which instead represents each scene as a combinatorial complex with five ranks of cells—nodes, edges, mesh triangles, collision contacts, and objects—and exchanges learnable messages along pathways chosen from Newtonian mechanics. On the MOVi-A and MOVi-B benchmarks, HOPNet reports lower position and orientation error than MeshGraphNet, FIGNet, and a same-split FIGNet reimplementation after 50 autoregressive rollout steps, and it stays below the best baseline's 50-step error for about 75 steps. The authors further show that a model trained only on spheres generalizes to complex and high-resolution unseen meshes, which they interpret as evidence that the model learns transferable face-to-face collision dynamics rather than object-specific patterns. If the claims hold, learned physics simulation should move from node-level graphs to higher-order topological representations with physically structured message passing.

What carries the argument

The load-bearing object is the spatiotemporal combinatorial complex $\mathcal{X}^t$, a hierarchy of cells of five ranks—nodes $\mathcal{X}^{(0)}$, edges $\mathcal{X}^{(1)}$, mesh triangles $\mathcal{X}^{(2)}$, collision contacts $\mathcal{X}^{(3)}$, and objects $\mathcal{X}^{(4)}$—each with its own learned features. The argument is carried by a sequential physics-informed message-passing scheme in the Encode-Process-Decode style: faces are first enriched with intra-object node, edge, and object messages; collisions are then processed at contact cells built from the closest points between triangles of different objects; and the collision effects are propagated back through objects to nodes before accelerations are decoded. The final pose comes from a second-order forward-Euler integration of node accelerations followed by shape matching, which enforces rigidity. This machinery is what lets the model treat collisions as surface-to-surface momentum exchanges rather than node-to-node events, and the ablations attribute the accuracy gains to exactly these components.

What would settle it

Retrain HOPNet and a faithful FIGNet reimplementation from the same codebase, data splits, random seeds, and compute budget, then compare 50-step RMSE on MOVi-A; if the margin disappears or reverses, the central claim fails. A second check: if HOPNet's translation RMSE at step 75 exceeds FIGNet's published 50-step value on MOVi-A, the claimed rollout-length advantage is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the information lost by graph representations—triangle-level surface structure, object-level cohesion, and explicit collision contacts—is exactly what a neural simulator needs to learn rigid-body collisions, and that a topological message-passing network can exploit it. HOPNet encodes the whole scene as a spatiotemporal combinatorial complex $\mathcal{X}^t$ with cells of ranks 0 through 4: nodes, edges, mesh triangles, collision contacts, and objects. Its message-passing proceeds in a fixed Newtonian order: enrich triangle embeddings with node, edge, and object information; compute collision effects between contacting triangles using the closest-point geometry of each face; propagate the collision updates back to objects and then to nodes; and decode per-node and per-object accelerations, which are integrated with second-order forward Euler and shape matching to produce the next pose. The paper reports that this design outperforms both graph baselines on MOVi-A and MOVi-B, that removing the object-level cells raises translation error by 44.4 percent, that removing the node-to-center-of-mass distance raises error by 26 percent on MOVi-A, and that replacing the sequential physics-guided message flow with unrestricted message passing raises error by 14 percent even with 34 percent more parameters.

Load-bearing premise

The comparison against FIGNet is the load-bearing premise: HOPNet's reported margin relies on the FIGNet numbers published in the original paper and on a reimplementation trained on the same splits, and if those baselines were produced under different conditions the advantage could shrink or disappear.

Editorial extensions

If this is right

On MOVi-A and MOVi-B after 50 rollout steps, HOPNet reports lower position and orientation RMSE than MeshGraphNet, FIGNet, and a same-split FIGNet reimplementation.
The model performs about 50 percent more rollout steps than the best baseline before reaching the same translation error, reaching FIGNet's 50-step RMSE only at step 75.
A model trained only on spheres generalizes to MOVi-A, MOVi-B, and the high-resolution scanned meshes of MOVi-C, and masking slow collisions during training improves generalization further.
Ablations show that object-level cells, the node-to-center-of-mass distance feature, and the sequential physics-guided message flow each contribute substantially to accuracy.
Because the representation is rebuilt at every timestep, the model can accommodate counterfactual interventions such as removing an object or changing initial conditions, and performance stays stable when the collision radius is halved or doubled.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same rank-structured message flow could be applied to deformable bodies or fluids by replacing rigid shape matching with a softer positional update, an extension the paper itself lists as future work.
Since the collision message is computed from local triangle geometry rather than object identity, a direct test with unseen object counts and mixed mesh resolutions would probe how far the universal collision dynamics claim extends.
Coupling HOPNet with a perception module that estimates mass, friction, and restitution from images would remove the current dependence on explicit state inputs and make the approach usable on raw video.
The small loss from cutting 74 percent of the parameters suggests representational structure, not raw capacity, drives performance, so hierarchical or adaptive mesh message passing may extend accuracy before model size becomes the bottleneck.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

HOPNet is a real contribution: the first (as far as I know) use of combinatorial complex message passing for rigid-body dynamics, with distinct cell types for triangles, collision contacts, and objects, and a sequential message schedule that mirrors collision-then-propagation. The ablations are clean—removing object-level cells, center-of-mass distance, and sequential scheduling all degrade performance as expected—and the authors ship code and data, which makes the work reproducible in principle. The OOD experiments, especially training on spheres and transferring to MOVi-C, are genuinely informative.

The soft spot is one the authors themselves expose: the SOTA comparison to FIGNet rests partly on numbers from the original paper and partly on a reimplementation described only in the supplement. That is a common situation when no official code exists, and the reimplementation baseline weakens the main concern, but it does not eliminate it. I would want to see the supplement's Section 2.5 and confirm that FIGNet_reimpl was given the same object-level features (mass, friction, restitution) as HOPNet before trusting the margin in Figure 4. If that check passes, the central claim likely holds. A smaller phrasing issue: "ensures energy conservation" in Methods 5.2.2 overstates what permutation-invariant learned messages do; they only bias the model toward symmetric updates.

None of this is fatal. The architecture is sensible, the evaluation is thorough, and the limitations (scaling to millions of faces, reliance on explicit state) are stated plainly. This is a paper for people working on learned physics simulation or topological deep learning; it will be useful to them as a baseline and a design source. I would send it to a serious referee: the review should be conditional on the FIGNet comparison being pinned down, but the work itself deserves the time.

Referee Report

3 major / 6 minor

Summary. The manuscript introduces HOPNet, a topological message-passing network for learning rigid-body dynamics from mesh representations. It represents an environment as a combinatorial complex whose cells correspond to nodes, edges, mesh triangles, collision contacts, and objects, and it orders message passing in a sequence motivated by Newtonian mechanics (collisions first, then object-level updates). Experiments on the Kubric MOVi-spheres/A/B/C datasets report lower position and orientation RMSE after 50 rollout steps than MeshGraphNet and FIGNet, improved long-horizon stability, OOD generalization to unseen mesh geometries, counterfactual rollouts, and ablation studies of the proposed components.

Significance. If the reported gains are confirmed under a fully controlled protocol, HOPNet would be a useful advance in learned rigid-body simulation on mesh inputs: it is one of the first demonstrations that higher-order/combinatorial complex representations combined with collision-ordered message passing can beat graph-network baselines on multi-object benchmarks. Strengths of the paper include public code/data, error bars over three seeds for HOPNet, explicit OOD evaluation, and ablation studies isolating the topological cells, the center-of-mass feature, and the sequential message-passing order. The main open issue is the fairness and completeness of the FIGNet comparison.

major comments (3)

[Section 3.1, Figure 4] The central comparison against MeshGraphNet and FIGNet is not fully controlled. The figure combines numbers taken directly from [6] with results from a reimplementation (FIGNet_reimpl) whose details are only in Supplementary Section 2.5; the published numbers may have been obtained under different data splits, object scaling, or evaluation protocols, and no uncertainty is reported for them. Because the abstract's claim of 'superior accuracy' is based on this comparison, the authors should either reproduce all baselines under the exact conditions used for HOPNet or explicitly restrict the state-of-the-art claim to the controlled reimplementation comparison.
[Section 3.1, Eq. (6)] The paper does not state whether the FIGNet_reimpl baseline received the same privileged physical parameters that HOPNet uses: mass m, friction coefficient c1, and restitution coefficient c2 in the object features h^(4) (Eq. 6). If FIGNet_reimpl did not have access to these parameters, the comparison is not architecture-vs-architecture. Please report the full feature set of the reimplementation and, if it differs, run a variant with matched inputs.
[Section 3.1, paragraph beginning 'Importantly'] The quantitative claim that HOPNet performs '50% more rollout steps' before reaching the same error as the best baseline is not supported by curves in the main text and again relies on the uncontrolled baseline numbers from [6]. The full error-versus-horizon curves for all methods under identical conditions should be shown, or the claim should be removed.

minor comments (6)

[Section 3.4, Figures 8 and 9] The text refers to 'Figure 8a', 'Figure 8b', and 'Figure 8c' when discussing the ablations; these should be 'Figure 9a', 'Figure 9b', and 'Figure 9c'.
[Eqs. (10)-(11) and surrounding text] The symbol m3->2_s->r is used in Eq. (11), but the text below the equation calls it m2->3_s->r; please make the notation consistent.
[Section 5.2.2] The phrase 'To ensure energy conservation' overstates what is implemented: no conservation constraint is enforced in the learned messages. Suggest rewording to 'motivated by' or 'designed to encourage'.
[Section 3.2] The phrase 'This counterintuitive results' should be 'This counterintuitive result'.
[Section 3.1] The phrase 'trained on exactly the same datasets than HOPNet' should be 'the same datasets as HOPNet'.
[Figure 4 caption] It would be clearer to state explicitly which bars come from [6] and which are from the reimplementation, since the legend includes both FIGNet and FIGNetreimpl.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HOPNet's claims are benchmark results for a learned model, not derivations from fitted inputs.

full rationale

Walking the paper's derivation chain, no step reduces by construction to its own inputs. HOPNet's outputs are learned predictions trained on Kubric-generated rollouts; the physics-informed message-passing pathways are architectural inductive biases rather than fitted quantities disguised as predictions, and the collision radius dc is a hyperparameter whose robustness is explicitly tested rather than tuned to match a target. The equations used (finite-difference features, message-passing updates, Euler integration, shape matching, and RMSE metrics) are standard computational definitions, and no equation is equivalent to the claimed result by construction. The headline SOTA comparison in Figure 4 uses external published numbers from FIGNet/MeshGraphNet plus a reimplementation; concerns about whether the baseline was faithfully reproduced or given the same physical features are experimental fairness issues, not circularity. The only self-citation is [24] (Han, Kammer, and Fink), cited in a general list of particle-based methods and not load-bearing for any central claim. There is no imported uniqueness theorem, no ansatz smuggled in via self-citation, and no renaming of a known result presented as unification. Accordingly, the paper is self-contained with respect to circularity concerns.

Assumptions & free parameters 1 free parameters · 5 assumptions · 2 invented entities

The ledger shows that the central empirical claims rest on supervised training over Kubric-generated synthetic data, on local proximity-based collision detection, on access to per-object mass, friction, and restitution, and on shape matching as a deterministic pose decoder. The only hand-set modeling threshold identified is the collision radius d_c; all other parameters are learned by gradient descent. No new physical entity with independent evidence is introduced: the higher-order cells are representational constructs whose value is supported only by the paper's own benchmarks.

free parameters (1)

Collision radius d_c = not stated; tested at 0.5x and 2x of default
Determines when two mesh triangles from different objects become a collision contact cell X(3). It is a hand-set threshold, although the paper reports low sensitivity to it.

assumptions (5)

domain assumption The MOVi/Kubric synthetic datasets provide physically accurate ground-truth states for rigid body dynamics.
The entire evaluation in Section 3.1 treats Kubric-generated videos as ground truth; if the simulator's contact model is wrong or simplified, the reported errors are measured against a questionable target.
ad hoc to paper Collisions are fully captured by proximity-based triangle pairs within radius d_c.
Section 5.1 creates X(3) cells when triangles of different objects are within d_c. This assumes local triangle proximity is sufficient to model contact, excluding cases such as deep interpenetration or long-range stacking forces.
domain assumption Shape matching recovers the true rigid pose from predicted node positions.
Section 5.2.4 uses shape matching to deterministically compute object positions and orientations from predicted node clouds; if the predicted nodes are noisy or partial, the pose fit may not correspond to the physical object.
ad hoc to paper The physics-informed message order, faces enriched, collisions processed, objects updated, is sufficient to represent collision dynamics.
The sequential schedule in Section 5.2 is an architectural assumption. The ablation in Section 3.4 supports it empirically, but there is no proof that this order captures all possible rigid body interactions.
domain assumption Mass, friction coefficient, and restitution coefficient of each object are known inputs at inference.
These features appear in Equation 6; if they are unknown, as the introduction's motivation suggests, HOPNet cannot use its stated feature set.

invented entities (2)

Higher-order topological cells X(2), X(3), X(4) for mesh triangles, collision contacts, and objects
purpose: Provide explicit message-passing units above the node level so that surfaces, contact events, and whole objects can exchange information.
These cells are a modeling construct invented for this architecture. Their usefulness is supported only by the paper's self-reported benchmark results, not by an independent falsifiable prediction.
Collision contact cells X(3)
purpose: Represent pairwise interactions between mesh triangles of different objects when they come within radius d_c.
These cells are created dynamically from proximity; they are not independently observable and are a design choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Integrating Physics and Topology in Neural Networks for Learning Rigid Body Dynamics." pith.science (2026). https://pith.science/paper/KCKQRRYG

@misc{pith2026241111467,
  author       = {Pith},
  title        = {Pith review of: Integrating Physics and Topology in Neural Networks for Learning Rigid Body Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KCKQRRYG}},
  note         = {Machine review of arXiv:2411.11467}
}

read the original abstract

Rigid body interactions are fundamental to numerous scientific disciplines, but remain challenging to simulate due to their abrupt nonlinear nature and sensitivity to complex, often unknown environmental factors. These challenges call for adaptable learning-based methods capable of capturing complex interactions beyond explicit physical models and simulations. While graph neural networks can handle simple scenarios, they struggle with complex scenes and long-term predictions. We introduce a novel framework for modeling rigid body dynamics and learning collision interactions, addressing key limitations of existing graph-based methods. Our approach extends the traditional representation of meshes by incorporating higher-order topology complexes, offering a physically consistent representation. Additionally, we propose a physics-informed message-passing neural architecture, embedding physical laws directly in the model. Our method demonstrates superior accuracy, even during long rollouts, and exhibits strong generalization to unseen scenarios. Importantly, this work addresses the challenge of multi-entity dynamic interactions, with applications spanning diverse scientific and engineering domains.

Figures

Figures reproduced from arXiv: 2411.11467 by the authors.

**Figure 2.** Spatiotemporal combinatorial complexes (a) Real-world observations are represented by (b) spatiotemporal combinatorial complexes {X t}. The topology and features of each combinatorial complex evolve over time to accurately represent the environment. The illustrated sequence is taken from experiment 1 of the MOVi-A dataset. To partially address those limitations, Allen et. al. [6] proposed duplicating shared object p… view at source ↗

**Figure 4.** Autoregressive rollout performance on benchmark datasets. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Rollout example on the MOVi-A dataset. (a) Ground truth1 ; (b) Model predictions. Our model makes very accurate and realistic predictions on the trajectories of rigid bodies despite complex multi-objects collisions and long rollout durations. The illustrated sequence i…

**Figure 8.** Figure 8: Counterfactual rollout example. (a) Original experiment2 ; (b) Altered experiment with one sphere removed, shown as a black skeleton to ease understanding. With our dynamic and flexible higher-order topological representation of the environment, our model can easily ac…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

HOPSE: Scalable Higher-Order Positional and Structural Encoder for Combinatorial Representations
cs.LG 2025-05 conditional novelty 6.0 of 10

HOPSE encodes higher-order topological data by applying graph positional and structural encoders to Hasse graph decompositions, matching or exceeding message-passing models on benchmarks with up to 7x faster training.

Reference graph

Works this paper leans on

50 extracted references · 43 canonical work pages · cited by 1 Pith paper

[6]

Allen, K. R. et al. Learning rigid dynamics with face interaction graph networks. InInternational Conference on Learning Representations (ICLR) (2023)

work page 2023
[1]

& Tassa, Y

Todorov, E., Erez, T. & Tassa, Y . Mujoco: A physics engine for model-based control. In2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , 5026–5033 (2012)

work page 2012
[2]

Bullet physics simulation

Coumans, E. Bullet physics simulation. In ACM SIGGRAPH 2015 Courses, SIGGRAPH ’15 (Association for Computing Machinery, New York, NY , USA, 2015)

work page 2015
[3]

Open Dynamics Engine (2005)

Smith, R. Open Dynamics Engine (2005). URL https://www.ode.org/

work page 2005
[4]

Nvidia PhysX system software (2008)

Corporation, N. Nvidia PhysX system software (2008). URL https://github.com/ NVIDIA-Omniverse/PhysX

work page 2008
[5]

& Battaglia, P

Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A. & Battaglia, P. W. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations (ICLR) (2021)

work page 2021
[7]

& Posa, M

Parmar, M., Halm, M. & Posa, M. Fundamental challenges in deep learning for stiff contact dynamics. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 5181–5188 (2021)

work page 2021
[8]

S., Weinstein, A., Battaglia, P

Piloto, L. S., Weinstein, A., Battaglia, P. & Botvinick, M. Intuitive physics learning in a deep- learning model inspired by developmental psychology. Nature Human Behaviour 6, 1257–1267 (2022)

work page 2022

Show all 50 references

[9]

Kipf, T. et al. Conditional Object-Centric Learning from Video. In International Conference on Learning Representations (ICLR) (OpenReview.net, 2022). URL https://openreview. net/forum?id=aD7uesX1GF_

2022
[10]

Elsayed, G. et al. Savi++: Towards end-to-end object-centric learning from real-world videos. In Koyejo, S. et al. (eds.) Advances in Neural Information Processing Systems , vol. 35, 28940– 28954 (Curran Associates, Inc., 2022)

2022
[11]

& Ahn, S

Singh, G., Wu, Y .-F. & Ahn, S. Simple unsupervised object-centric learning for complex and naturalistic videos. In Koyejo, S. et al. (eds.) Advances in Neural Information Processing Systems, vol. 35, 18181–18196 (Curran Associates, Inc., 2022). 17

2022
[12]

& Garg, A

Wu, Z., Dvornik, N., Greff, K., Kipf, T. & Garg, A. Slotformer: Unsupervised visual dynamics simulation with object-centric models. In International Conference on Learning Representations (ICLR) (2023)

2023
[13]

Burgess, C. P. et al. MONet: Unsupervised scene decomposition and representation (2019). URL https://arxiv.org/abs/1901.11390. 1901.11390

2019 arXiv
[14]

Locatello, F. et al. Object-centric learning with slot attention. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M. & Lin, H. (eds.) Advances in Neural Information Processing Systems , vol. 33, 11525–11538 (Curran Associates, Inc., 2020)

2020
[15]

Vaswani, A. et al. Attention is all you need. In Guyon, I. et al. (eds.) Advances in Neural Information Processing Systems, vol. 30 (Curran Associates, Inc., 2017)

2017
[16]

& Botvinick, M

Ding, D., Hill, F., Santoro, A., Reynolds, M. & Botvinick, M. Attention over learned object embeddings enables complex visual reasoning. In Ranzato, M., Beygelzimer, A., Dauphin, Y ., Liang, P. & Vaughan, J. W. (eds.)Advances in Neural Information Processing Systems , vol. 34,...

2021
[17]

Chen, Z. et al. Comphy: Compositional physical reasoning of objects and events from videos. In International Conference on Learning Representations (ICLR) (2022)

2022
[18]

Bear, D. et al. Physion: Evaluating physical prediction from vision in humans and machines. In Vanschoren, J. & Yeung, S. (eds.)Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, vol. 1 (2021)

2021
[19]

C., Hagenbuchner, M

Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M. & Monfardini, G. The graph neural network model. IEEE transactions on neural networks 20, 61–80 (2008)

2008
[20]

Mrowca, D. et al. Flexible neural representation for physics prediction. In Bengio, S. et al. (eds.) Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada , 8813...

2018
[21]

Propagation networks for model-based control under partial observation

Li, Y .et al. Propagation networks for model-based control under partial observation. In 2019 International Conference on Robotics and Automation (ICRA) , 1205–1211 (IEEE, 2019)

2019
[22]

Li, Y ., Wu, J., Tedrake, R., Tenenbaum, J. B. & Torralba, A. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. In ICLR (2019)

2019
[23]

Sanchez-Gonzalez, A. et al. Learning to simulate complex physics with graph networks. In III, H. D. & Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning, vol. 119 of Proceedings of Machine Learning Research, 8459–8468 (PMLR, 2020)

2020
[24]

Han, Z., Kammer, D. S. & Fink, O. Learning physics-consistent particle interactions. PNAS Nexus 1, pgac264 (2022)

2022
[25]

Han, J. et al. Learning physical dynamics with subequivariant graph neural networks. In Koyejo, S. et al. (eds.) Advances in Neural Information Processing Systems , vol. 35, 26256–26268 (Curran Associates, Inc., 2022)

2022
[26]

Allen, K. R. et al. Graph network simulators can learn discontinuous, rigid contact dynamics. In Liu, K., Kulic, D. & Ichnowski, J. (eds.) Proceedings of The 6th Conference on Robot Learning , vol. 205 of Proceedings of Machine Learning Research, 1157–1167 (PMLR, 2023)

2023
[27]

Battaglia, P. W.et al. Relational inductive biases, deep learning, and graph networks (2018). URL http://arxiv.org/abs/1806.01261. 1806.01261

2018 arXiv
[28]

& Suzuki, T

Oono, K. & Suzuki, T. Graph Neural Networks Exponentially Lose Expressive Power for Node Classification. In International Conference on Learning Representations (ICLR) (2020)

2020
[29]

Battiston, F. et al. The physics of higher-order interactions in complex systems. Nature Physics 17, 1093–1098 (2021). 18

2021
[30]

& Battiston, F

Zhang, Y ., Lucas, M. & Battiston, F. Higher-order interactions shape collective dynamics differently in hypergraphs and simplicial complexes. Nature Communications 14, 1605 (2023)

2023
[31]

Malizia, F. et al. Reconstructing higher-order interactions in coupled dynamical systems. Nature Communications 15, 5184 (2024)

2024
[32]

T., Sauceda, H

Schütt, K. T., Sauceda, H. E., Kindermans, P.-J., Tkatchenko, A. & Müller, K.-R. SchNet – A deep learning architecture for molecules and materials. The Journal of Chemical Physics 148, 241722 (2018)

2018
[33]

Fang, X. et al. Geometry-enhanced molecular representation learning for property prediction. Nature Machine Intelligence 4, 127–134 (2022)

2022
[34]

& Wei, G.-W

Cang, Z. & Wei, G.-W. Topologynet: Topology based deep convolutional and multi-task neural networks for biomolecular property predictions. PLoS computational biology 13, e1005690 (2017)

2017
[35]

& Gastegger, M

Schütt, K., Unke, O. & Gastegger, M. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In Proceedings of the 38th International Conference on Machine Learning, 9377–9388 (PMLR, 2021)

2021
[36]

Pineda, J. et al. Geometric deep learning reveals the spatiotemporal features of microscopic motion. Nature Machine Intelligence 5, 71–82 (2023)

2023
[37]

Hajij, M. et al. Topological deep learning: Going beyond graph data. URL https://arxiv. org/abs/2206.00606. 2206.00606

arXiv
[38]

Topological Deep Learning: Graphs, Complexes, Sheaves

Bodnar, C. Topological Deep Learning: Graphs, Complexes, Sheaves . Ph.D. thesis, Apollo - University of Cambridge Repository (2022). URL https://doi.org/10.17863/CAM. 97212

2022 doi
[39]

& Miolane, N

Papillon, M., Sanborn, S., Hajij, M. & Miolane, N. Architectures of topological deep learning: A survey of message-passing topological neural networks (2024). URL https://arxiv.org/ abs/2304.10031. 2304.10031

2024 arXiv
[40]

M., Glaze, N

Roddenberry, T. M., Glaze, N. & Segarra, S. Principled simplicial neural networks for trajectory prediction. In Meila, M. & Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning , vol. 139 of Proceedings of Machine Learning Research , 9020–9029 ...

2021
[41]

Bodnar, C. et al. Weisfeiler and Lehman go cellular: CW networks. In Ranzato, M., Beygelzimer, A., Dauphin, Y ., Liang, P. & Vaughan, J. W. (eds.)Advances in Neural Information Processing Systems, vol. 34, 2625–2640 (Curran Associates, Inc., 2021)

2021
[42]

& Gross, M

Müller, M., Heidelberger, B., Teschner, M. & Gross, M. Meshless deformations based on shape matching. ACM Trans. Graph. 24, 471–478 (2005)

2005
[43]

Greff, K. et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 3749–3761 (2022)

2022
[44]

Downs, L. et al. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA) , 2553–2560 (IEEE, 2022)

2022
[45]

Battaglia, P., Pascanu, R., Lai, M., Jimenez Rezende, D. et al. Interaction networks for learning about objects, relations and physics. Advances in neural information processing systems 29 (2016)

2016
[46]

& Heckbert, P

Garland, M. & Heckbert, P. S. Surface simplification using quadric error metrics. In Owen, G. S., Whitted, T. & Mones-Hattal, B. (eds.) Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1997, Los Angeles, CA, USA, August 3-8, 1...

1997
[47]

& O’brien, J

Narain, R., Samii, A. & O’brien, J. F. Adaptive anisotropic remeshing for cloth simulation. ACM transactions on graphics (TOG) 31, 1–10 (2012). 19

2012
[48]

Bleeker, M. et al. Neuralcfd: Deep learning on high-fidelity automotive aerodynamics simula- tions. arXiv preprint arXiv:2502.09692 (2025)

2025
[49]

Paszke, A. et al. Pytorch: An imperative style, high-performance deep learning library. In Wallach, H. et al. (eds.) Advances in Neural Information Processing Systems , vol. 32 (Curran Associates, Inc., 2019)

2019
[50]

Topox: A suite of python packages for machine learning on topological domains (2024)

PYT-Team. Topox: A suite of python packages for machine learning on topological domains (2024). URL https://arxiv.org/abs/2402.02441. 2402.02441. Acknowledgments We thank Vinay Sharma and Ismail Nejjar for helpful discussions. This work was supported by the Swiss National Scie...

2024 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [6]

Allen, K. R. et al. Learning rigid dynamics with face interaction graph networks. InInternational Conference on Learning Representations (ICLR) (2023)

work page 2023

[2] [1]

& Tassa, Y

Todorov, E., Erez, T. & Tassa, Y . Mujoco: A physics engine for model-based control. In2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , 5026–5033 (2012)

work page 2012

[3] [2]

Bullet physics simulation

Coumans, E. Bullet physics simulation. In ACM SIGGRAPH 2015 Courses, SIGGRAPH ’15 (Association for Computing Machinery, New York, NY , USA, 2015)

work page 2015

[4] [3]

Open Dynamics Engine (2005)

Smith, R. Open Dynamics Engine (2005). URL https://www.ode.org/

work page 2005

[5] [4]

Nvidia PhysX system software (2008)

Corporation, N. Nvidia PhysX system software (2008). URL https://github.com/ NVIDIA-Omniverse/PhysX

work page 2008

[6] [5]

& Battaglia, P

Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A. & Battaglia, P. W. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations (ICLR) (2021)

work page 2021

[7] [7]

& Posa, M

Parmar, M., Halm, M. & Posa, M. Fundamental challenges in deep learning for stiff contact dynamics. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 5181–5188 (2021)

work page 2021

[8] [8]

S., Weinstein, A., Battaglia, P

Piloto, L. S., Weinstein, A., Battaglia, P. & Botvinick, M. Intuitive physics learning in a deep- learning model inspired by developmental psychology. Nature Human Behaviour 6, 1257–1267 (2022)

work page 2022