Our Vision
Move from AI that only predicts toward AI that discovers, verifies, and learns within the boundaries of engineering physics.
Accurate predictions are important. Physically credible predictions are essential.
Artificial intelligence can accelerate engineering calculations, but a model with low prediction error does not necessarily satisfy physical laws.
Our research on Physics-Audited Agentic Scientific Machine Learning addresses this fundamental challenge.
The framework combines autonomous AI-assisted model discovery with explicit engineering verification. Specialized agents explore, implement, train, and evaluate candidate machine-learning models, while numerical physics audits assess their compliance with defined physical requirements.
Unlike conventional approaches focused primarily on minimizing prediction error, the proposed work considers both predictive performance and evidence of physical admissibility.
The framework is designed to:
Move from AI that only predicts toward AI that discovers, verifies, and learns within the boundaries of engineering physics.
Demonstrations in computational mechanics illustrate how physics auditing can distinguish models with apparently strong predictive accuracy from models that also satisfy the specified physical checks.
This research provides a common AI foundation for our future Structural and Soil Digital Twin initiatives, including intelligent model discovery, uncertainty quantification, and reliability assessment.