Advancing Surrogate Modeling Techniques for Efficient Uncertainty Quantification in Natural Hazards Engineering

Advancements in computing power and modeling capabilities are enabling the assessment of natural hazards and their impact on the built environment at an unprecedented scale and resolution. However, these high-fidelity, large-scale simulations also present new challenges to the field of computational Uncertainty Quantification (UQ), driving innovations in computational statistics and machine learning. One example is the development of surrogate modeling techniques, which allow us to approximate the outcomes of complex simulations using faster statistical methods. 
This talk will highlight recent advancements in Gaussian process surrogate modeling, developed to address practical challenges encountered in natural hazards engineering. These challenges include high-dimensional outputs, the presence of aleatoric uncertainty and its correlations, and the need for adaptive training set selection to minimize computational burdens. Additionally, the talk will explore how surrogate modeling can be integrated into broader many-query workflows, such as Bayesian calibration, design optimization, and structural risk and resilience assessment, to reduce computational demands while maintaining the desired fidelity.