Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation

TL;DR

This paper investigates the use of latent diffusion models for fast physics emulation, demonstrating robustness to compression and improved accuracy over non-generative models, with insights into effective training practices.

Contribution

It is the first comprehensive empirical study applying latent diffusion models to physics emulation, highlighting their robustness and practical training considerations.

Findings

Latent-space emulation remains accurate at high compression rates.

Diffusion-based emulators outperform non-generative models in accuracy.

Latent diffusion models provide greater diversity to compensate for uncertainty.

Abstract

The steep computational cost of diffusion models at inference hinders their use as fast physics emulators. In the context of image and video generation, this computational drawback has been addressed by generating in the latent space of an autoencoder instead of the pixel space. In this work, we investigate whether a similar strategy can be effectively applied to the emulation of dynamical systems and at what cost. We find that the accuracy of latent-space emulation is surprisingly robust to a wide range of compression rates (up to 1000x). We also show that diffusion-based emulators are consistently more accurate than non-generative counterparts and compensate for uncertainty in their predictions with greater diversity. Finally, we cover practical design choices, spanning from architectures to optimizers, that we found critical to train latent-space emulators.