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

Area of Science:

• Structural Engineering
• Artificial Intelligence
• Computational Mechanics

Background:

• Accurate prediction of nonlinear structural dynamics is crucial for building design and disaster mitigation.
• Conventional finite-element analysis is computationally intensive, limiting real-time and large-scale applications.

Purpose of the Study:

• To develop SeisGPT, a physics-informed foundation model for high-fidelity, rapid response prediction in multistorey buildings.
• To overcome the computational bottleneck of traditional methods for structural dynamics analysis.

Main Methods:

• Developed SeisGPT, integrating structural mechanics priors, storey-resolved representations, and a spectral decoder.
• Trained the model on over 2 million nonlinear simulations from diverse structural designs and real-world buildings.
• Incorporated physics-informed learning and a spectral decoder for modal response propagation and nonlinear corrections.

Main Results:

• SeisGPT achieved normalized errors below 5% for predicting displacement and acceleration histories in unseen buildings.
• Demonstrated a computational speedup of approximately 40,000 times compared to finite-element analysis.
• Successfully reconstructed full-building responses from sparse sensor data and showed high accuracy in shake-table tests.

Conclusions:

• Physics-informed foundation models like SeisGPT offer a viable solution for high-fidelity structural response prediction.
• The model's speed and accuracy support real-time structural health monitoring and large-scale disaster resilience assessment.
• SeisGPT enables efficient structural assessment and damage evaluation, crucial for post-earthquake scenarios.