Collaborative Research: Reliable Physics-Guided Generative Artificial Intelligence for Constitutive Modeling of Fractured Rock Masses

This award supports research to improve the prediction of deformation and failure in fractured rock masses that host tunnels, underground storage facilities, energy systems, transportation infrastructure, foundations, and slopes. Existing computational models have limited ability to capture the coupled effects of interacting fracture networks, fluids, loading history, material heterogeneity, and progressive damage, often resulting in either unexpected failures or overly conservative and costly designs. The project will develop a computational framework that integrates physics-based modeling with data-driven learning to improve predictive accuracy while identifying the conditions under which model predictions can be trusted. The resulting approach will strengthen engineering analysis and provide more reliable tools for designing and managing safer, more economical, and more resilient critical infrastructure in the United States. Publicly available datasets, open-source software, documentation, and educational examples will promote transparency, independent validation, and continued innovation. The project will also provide interdisciplinary training for graduate and undergraduate students in geomechanics, computational modeling, and data science. Complementary outreach and classroom activities will expand participation and disseminate the research outcomes. The project will develop and evaluate a physics-guided generative adversarial network as a solver-ready constitutive stress-update operator for monotonic loading of single-layer fractured rock masses represented as equivalent continua. The model will be trained and validated using laboratory stress-strain data, field observations, and physics-consistent synthetic responses generated through validated implicit joint-continuum simulations. Conditioned on the evolving stress state, loading history, intact-rock properties, and statistical descriptors of fracture geometry and mechanics, the network will enforce conservation laws, thermodynamic admissibility, and nonnegative dissipation. Embedded at material integration points within a finite-element framework, the learned operator will return stress increments and consistent tangent moduli for robust nonlinear analysis. Regularization with an intrinsic length scale will address localization and mesh dependence. Verification will range from material-point loading paths to representative geotechnical boundary-value problems, including foundations, excavations, and underground cavities. The framework will incorporate uncertainty quantification, out-of-distribution detection, admissibility projection, adaptive step control, and fallback constitutive updates to ensure numerical robustness. The project will establish quantitative criteria for trustworthy physics-guided machine learning and deliver benchmark datasets, verification tests, open-source software, and comprehensive documentation for the geomechanics community. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.