PhysiX: A Foundation Model for Physics Simulations

Tung Nguyen, Arsh Koneru, Shufan Li, Aditya Grover

Keywords: Physics Simulation, PDEs, Foundation Models, Generative Models

TL;DR: We propose a scalable and performant foundation model for physics simulation

Abstract: While foundation models have achieved remarkable success in domains like video, image, and language by scaling on massive datasets, this progress has not yet translated to physics simulation. A primary bottleneck is data scarcity: while millions of images, videos, and textual resources are readily available on the internet, the largest physics simulation datasets contain only tens of thousands of samples. This data limitation makes large models prone to overfitting and has confined physics applications to small models, which struggle with complex domains and long-range predictions. Furthermore, the drastic variations in scale and structure across physics datasets—a heterogeneity not typically found in vision or language—further amplify the challenges of scaling up multitask training. We introduce PhysiX, a family of large-scale foundation models for physics simulation. PhysiX is an autoregressive generative model composed of a discrete tokenizer, which converts heterogeneous physical processes to sequences of tokens, and a Transformer that models these sequences via next-token prediction. To mitigate the rounding error in the discretization process, PhysiX additionally incorporates a specialized refinement module. Extensive experiments on 2D datasets in The Well benchmark show that PhysiX achieves superior performance over existing foundation models and strong task-specific baselines. Our results demonstrate that PhysiX benefits from synergistic learning through joint training on diverse simulation tasks and can successfully transfer knowledge from natural videos to the physical domain. We further analyze PhysiX’s generalization to unseen domains and conduct careful ablation studies to validate the impact of each design component.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 14058