Abstract: Many physical systems can end up in several different, equally valid states under exactly the same conditions. A compressed beam can buckle in any direction, and a mixture can separate into many different patterns. Such bifurcations occur throughout physics, but they are a blind spot for most machine-learning models. These models are built to predict one answer, so they tend to average over the possibilities and return a state that can’t physically exist. This talk presents Bi-FORK, a generative approach that learns the full range of possible outcomes instead of a single prediction. Earlier methods of this kind were limited to systems with a few hundred degrees of freedom. This approach is demonstrated on buckling structures and phase separation with up to 260,000 points. The talk also discusses what it takes to make machine learning trustworthy for systems where uniqueness can’t be taken for granted.
Date
Chair
Location
Speaker
Affiliation
Go to the Events page.