A physics-informed foundation model for rapid high-fidelity structural response prediction.

Abstract

Accurate and rapid prediction of nonlinear structural dynamics is central to structural design, disaster mitigation and resilience assessment, yet conventional finite-element analysis remains too computationally demanding for real-time or portfolio-scale applications. To address this computational bottleneck, we develop SeisGPT, a physics-informed foundation model for high-fidelity response prediction across multistorey buildings with diverse structural types and topologies. Its architecture integrates structural-mechanics priors, storey-resolved structural representations and a spectral decoder that captures modal response propagation while accommodating learnable nonlinear corrections. The model is trained on more than 2 million nonlinear elastoplastic simulations derived from 270,000 automatically generated, code-compliant structural designs and 694 real-world buildings, together comprising over 10 billion response time steps. For previously unseen buildings, SeisGPT predicts displacement and acceleration histories with normalized errors below 5% and achieves an approximately 40,000-fold computational speedup relative to conventional finite-element analysis. By assimilating sparse sensor measurements, SeisGPT reconstructs full-building responses from limited observations and, in shake-table tests, shows closer agreement with measured responses than finite-element simulations, supporting structural-health monitoring and post-earthquake damage-state assessment. These results indicate that physics-informed foundation models can support high-fidelity structural-response prediction, sparse-sensor reconstruction and response-informed downstream structural assessment at inference speeds compatible with real-time and portfolio-scale use.