Multiscale AI @ ICLR 2025

@multiscaleai.bsky.social

57 Followers 22 Following

ICLR 2025 workshop working building AI to answer a single question: Given low-level theory and computationally-expensive simulation code, how can we model complex systems on a useful time scale?

Standard graph neural networks assign vastly different latent embeddings to graphs describing the same physical system at different resolution scales. This precludes consistency in applications and prevents generalization between scales as would fundamentally be needed in many scientific applications. We uncover the underlying obstruction, investigate its origin and show how to overcome it.

Equivariant neural networks can effectively model physical systems by naturally handling the underlying geometric quantities and preserving their symmetries, but scaling them to large geometric data remains challenging. Naive downsampling typically disrupts features’ transformation laws, limiting their applicability in large scale settings. In this work, we propose a scalable equivariant transformer that efficiently processes geometric data in a coarse-grained latent space while preserving E(3) symmetries of the problem. In particular, by building on the Geometric Algebra Transformer (GATr) and PerceiverIO architectures, our method learns equivariant latent tokens which allow us to decouple the processing complexity from the input data representation while maintaining global equivariance.