Speaker
Description
Surrogate models such as PDE learners and machine learning force fields have become essential tools for approximating complex physical systems at a fraction of the computational cost of high-fidelity solvers. As these models are increasingly deployed in scientific and engineering workflows, quantifying their predictive uncertainty is critical for ensuring reliability and informing data acquisition strategies. Yet a persistent challenge remains: classical Bayesian inference methods, while principled, are often prohibitively expensive for the large-scale architectures that make modern surrogates so powerful. We present recent advances in Bayesian inference that treat computational efficiency as a first-class concern alongside trustworthiness and calibration. Our framework converts any surrogate model into an uncertainty-aware counterpart through a variational formulation that scales gracefully with model size and data volume, yielding calibrated uncertainty estimates without sacrificing the architectural flexibility that practitioners depend on. We further demonstrate how the framework naturally supports active learning, where predictive uncertainty drives adaptive sampling and experimental design, enabling more data-efficient construction of reliable surrogates.