We are pleased to announce that our paper, titled “Rapid training of Hamiltonian graph networks using random features” by Atamert Rahma, Chinmay Datar, Ana Cukarska, and Felix Dietrich has been accepted to the International Conference on Learning Representations (ICLR) 2026. The paper presents “Random Feature Hamiltonian Graph Networks”, where we demonstrate that Hamiltonian Graph Networks can be trained up to 150-600x faster - but comparable accuracy - by replacing iterative gradient-descent optimization with random feature based parameter construction. Our approach addresses the following challenges:
- Random Feature Hamiltonian Graph Networks are introduced, combining random sampling with graph-based physics-informed models for the first time.
- We avoid the challenges posed by slow, non-convex optimization in graph networks.
- We provide a comprehensive comparison against 15 different optimizers, including Adam and LBFGS.
- We demonstrate zero-shot generalization with models trained on tens of nodes accurately predicting dynamics with thousands of nodes.
arXiv: arxiv.org/abs/2506.06558