BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:UW-Madison-Physics-Events
BEGIN:VEVENT
SEQUENCE:2
UID:UW-Physics-Event-8389
DTSTART:20231109T160000Z
DTEND:20231110T000000Z
DTSTAMP:20260414T044419Z
LAST-MODIFIED:20231107T171410Z
LOCATION:5310 Chamberlin
SUMMARY:Scaling Deep Learning for Materials Discovery\, R. G. Herb Con
 densed Matter Seminar\, Ekin Dogus Cubuk\, Google Brain
DESCRIPTION:Despite the recent advances in physical simulations and ma
 chine learning\, the exploration of novel inorganic crystals remains c
 onstrained by the expensive trial-and-error approaches. Recent develop
 ments in deep learning have shown that models can showcase emergent pr
 edictive capabilities with increasing data and computation in fields s
 uch as language\, vision\, and biology. In this talk\, I will present 
 our recent results on how graph networks trained at scale can reach un
 precedented levels of generalization\, improving the efficiency of mat
 erials discovery by an order of magnitude. Building on the 48\,000 sta
 ble crystals identified in ongoing studies\, improved efficiency enabl
 es the discovery of 2.2 million stable structures with respect to the 
 current convex hull\, many of which had escaped prior human chemical i
 ntuition. The scale and diversity unlock surprising modeling capabilit
 ies for downstream applications\, leading in particular to highly accu
 rate and robust learned interatomic potentials that can be used in con
 densed-phase molecular dynamics simulations and high-fidelity zero-sho
 t predictions.
URL:https://www.physics.wisc.edu/events/?id=8389
END:VEVENT
END:VCALENDAR