Multifidelity surrogates and decision-making for digital twins
Anirban Chaudhuri
UT Austin Oden Institute
Multifidelity methods provide a principled framework for combining hierarchies of models and data sources of varying cost and fidelity to enable scalable prediction and decision-making for complex systems. This talk focuses on surrogate modeling and decision-making methods that explicitly exploit such hierarchies. On the surrogate modeling side, multifidelity linear regression approaches based on combined loss functions are presented, along with extensions to nonlinear regression and reduced-order modeling. Multifidelity Gaussian-process–based active learning methods for targeted contour and failure boundary location are also discussed, as they play a central role in reliability analysis and stability analysis. On the decision-making side, strategies for uncertainty-aware decision-making are presented, including information reuse for importance sampling in risk-based design optimization and control-variates–based Monte Carlo estimators for robust optimization. Taken together, these methods can be framed within a multifidelity view of digital twins as uncertainty-aware, computationally efficient virtual constructs for prediction and risk-informed decision-making and are directly applicable to safety-critical domains such as aerospace, medicine, and fusion plasma physics.