Title: Toward a Foundational Framework for Physics

Abstract:

The "train once, deploy anywhere" methodology has fundamentally reshaped natural language processing, allowing a single pre-trained architecture to adapt to a vast array of downstream applications without the need for retraining. A comparable breakthrough in physics—access to a Physics Foundation Model (PFM)—would be revolutionary, potentially democratizing high-fidelity simulation capabilities, speeding up scientific inquiry, and removing the burden of developing specialized solvers for every new problem. However, existing physics-aware machine learning techniques are currently constrained to narrow, isolated domains and necessitate retraining whenever a new system is introduced.

In this work, we introduce the General Physics Transformer (GPhyT), a model trained on 1.8 TB of heterogeneous simulation data, which proves that foundation model paradigms are viable within the physical sciences. Our central hypothesis is that transformers can deduce governing dynamics from contextual information, allowing one unified model to simulate diverse phenomena—including fluid-solid interactions, shock waves, thermal convection, and multi-phase dynamics—without explicit instruction regarding the underlying equations. GPhyT delivers three major advancements: (1) it surpasses specialized architectural designs by a factor of more than 7x across various physics domains; (2) it exhibits plausible zero-shot generalization to completely novel physical systems via in-context learning; and (3) it provides more stable long-term predictions through extended rollouts. By demonstrating that a single model can extract generalizable physical laws directly from data, this research paves the way for a universal PFM capable of transforming computational science and engineering.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC