Teodora Srećković

The automated design of experiments for fundamental physics is an emerging topic in ML of high potential scientific impact. In contrast to "experimental design" in the classic sense, this setting features high-dimensional design spaces, with very narrow, non-convex manifolds of good solutions. And in contrast to generative tasks in current industrial ML, data is very scarce. However, simulations can be run at comparably low computational cost, and even provide gradients. Collaborating with local partners, we plan to explore opportunities for new ideas in Bayesian experimental design, leveraging diffusion models calibrated by nontrivial Gaussian priors, and recently emerging Bayesian optimization techniques for high-dimensional problems.