acceptodds
Under review as a conference paper at ICLR 2027

Multi-resolution energy diffusion graph neural network discovers system conservation variables

Abstract

Discovering conservation laws and assessing integrability are fundamental to understanding the evolution and long-term behavior of Hamiltonian dynamical systems. The neural deflation method provides a principled framework for iteratively learning conservation laws, but its performance depends on the representational capacity of the underlying neural network. Standard multilayer perceptrons lack structural priors and struggle to scale to high-dimensional lattice systems, leading to degraded accuracy and limited interpretability. We propose incorporating a multi-resolution diffusion graph neural network (MR-DGNN) as the approximator for conserved quantities within neural deflation. MR-DGNN uses multi-resolution graph diffusion operators to adaptively capture long-range and local particle interactions. This architecture aligns with the discrete geometric structure of lattice systems and achieves higher accuracy with significantly fewer parameters than baselines. Moreover, the learned diffusion weights reveal the spatial interaction range of each conserved quantity, supporting physical interpretation. We evaluate the method on several benchmark systems. Results show that MR-DGNN consistently outperforms baselines in test error, with advantages becoming more pronounced as the number of degrees of freedom increases. The interpretable weight patterns further indicate that different conserved quantities exhibit distinct interaction ranges. These findings suggest that the method can accurately discover conservation laws while providing physical insights into complex high-dimensional Hamiltonian systems.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph neural network, conservation laws, lattice Hamiltonian system

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ArsmOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MR-DGNN, a multi-resolution graph diffusion architecture integrated into neural deflation for discovering conservation laws in Hamiltonian systems. By exploiting spatial interactions across multiple scales, the method aims to improve accuracy and parameter efficiency while providing interpretable interaction patterns. Experiments on four benchmark systems report improvements over MLP, CNN, GCN, and DeepSets baselines, particularly as system dimensionality increases.

Critical strengths
  • Integrating multi-resolution graph diffusion with neural deflation provides a promising way to incorporate spatial structure into conservation law discovery.

  • The evaluation covers both integrable and non-integrable Hamiltonian systems, offering a reasonably diverse assessment of the proposed architecture.

  • The reported improvements in approximation accuracy and parameter efficiency are encouraging, particularly for systems with more degrees of freedom.

Critical weaknesses
  • In Section 4, Definition 2 sets $ L=I_N-A$ using an unnormalized adjacency matrix, contradicting the earlier normalized Laplacian formulation. For a degree-two lattice, Equation 1 consequently amplifies constant signals by a factor of 1+κ1+\kappa, rather than preserving them as a valid diffusion operator should.
  • In Section 4.1 and Figure 3, the authors associate H3H_3 with kinetic energy, although kinetic energy alone is generally not conserved in the Toda lattice. Moreover, aggregated diffusion weights do not directly characterize the learned conservation laws because their contributions are subsequently transformed by nonlinear networks.
Questions for authors

see wekaness.

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

This paper suggests relying on a graph neural diffusion-inspired GNN inside the established neural deflation method for learning conserved quantities within a multi-particle system. It proposes a GNN with layer-wise step sizes for the numerical solution, hence claiming multi-resolution. The experimental suite compares this model on four small-scale systems to a set of simple baselines, with the main claim to outperform a simple MLP, which seems to hold statistically when scaling the systems to 32 particles.

Critical strengths

Learning conservation quantities from observations is a relevant research field, especially leveraging spatial-aware models like graph neural networks.

Critical weaknesses
  • the multi-resolution claim is overly strong, given the simple time-discretization flexibility. There is no real multi-resolution in space, and no ablation is provided to support the empirical benefit over a constant step size.
  • no details about the graph construction are provided, and no obvious spatial ordering for a CNN is given
  • the post-training investigation of spatial dependence of each conservation law is not connected with any ground-truth source of actual spatially dependent conserved quantities and uses an arbitrary-looking number of five spatial aggregation layers.
  • no comparison to Hamiltonian-inspired graph neural networks or existing graph neural diffusion models in the recent literature and statistically weak gains over the baselines
Questions for authors
  • Please clarify the ground truth of a spatial-dependend conserved quantitiy to justify the post-training analysis
  • Please could you provide a large-scale experiment going beyond 32 particles?
  • Please could you present a related work section on the multi-resultion method you are developing?
  • Please could you provide additional experimental data justifying the number of layers and how the train,val and test splits are generated on the data-generating systems?
  • Please could you elaborate on how you arranged the particles for a GNN and CNN to be used?
  • Please could you further explain how the supervised training loss in Section 3 can be materialized with learning only on observations?
Reviewer_iMSeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper replaces the MLP module in the nerual deflation method by the GNN based diffusion module, which helps to better capture the intrinsic interaction of the particles in the Hamiltonian systems. Experiments on four Hamiltonian systems report improved accuracy and parameter efficiency over several neural network baselines.

Critical strengths
  1. The graph architecture provides a suitable inductive bias for many body systems.

  2. The reported results show improved accuracy with fewer parameters than the MLP baseline.

Critical weaknesses
  1. The main contribution is replacing its the MLP modules in the previous method with a multi-resolution graph network, which limits the methodological novelty.

  2. The experiments do not assess noisy observations or errors in vector-field estimates. Since the objective requires vector-field evaluations, its reliability with measured trajectories remains unclear.

  3. The baselines mainly compare network architectures within neural deflation method and GNN based methods. The paper does not establish why the methods of Liu and Tegmark (2021), Liu et al. (2022, 2024), or Lu, Dangovski, and Soljačić (2023) are unsuitable for these benchmarks. The comparison therefore does not sufficiently position the method against alternative discovery frameworks.

Inconsistent Laplacian definition. Section 4 first defines WW as a random-walk matrix, but later sets L=I−AL=I-A for an unweighted adjacency matrix. Unit edge weights do not imply row normalization. This discrepancy affects the interpretation of the proposed operator as diffusion.

Questions for authors
  1. How does performance change under state noise and inaccurate vector-field estimates?

  2. Do you require the dataset to be regular time series?

  3. What prevents the methods of Liu and Tegmark (2021) and Liu et al. (2022, 2024) from being applied to these benchmarks?

  4. Could the authors compare against mainstream methods such as Lu et al. (2023), Discovering conservation laws using optimal transport and manifold learning, or explain the relevant differences in data requirements and applicability?

  5. In Definition 2, does L=IN−AL=I_N-A satisfy the normalisation of Laplacian operator?

  6. In Line 217, why the dimension of xx is N×dN\times d instead of N×dN\times d, does dd refer to the dimension of position state qq?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.