Advances in Bayesian optimization for accelerating scientific decision-making

With breakthroughs in scientific computing, virtual simulators are increasingly used as "digital twins" for studying complex scientific phenomena, e.g., particle collisions and nuclear reactions. Such simulators, however, are highly time-intensive, which hinders their effective use for scientific decision-making, e.g., the optimization of a particle detector. There is a pressing need for novel Bayesian ML/AI methods that synergize with digital twins to enable timely, uncertainty-aware, and interpretable decision-making. One promising tool is Bayesian optimization (BO), which tackles the optimization of a costly black-box function (e.g., the response surface of the virtual simulator) using limited function evaluations. In this talk, I will present a suite of novel BO methods that address several important needs for scientific applications. The first is a new BO method for identifying varied local optima of a black-box function, which provides scientists with a basket of different optimal solutions for flexible decision-making. The second method tackles the challenging problem of high-dimensional black-box optimization, offering improved theoretical rates and empirical performance over existing techniques in the one-shot setting. The effectiveness of these methods will be investigated in ongoing collaborative projects in the physical and engineering sciences.