BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:"Machine learning in chemistry and beyond" (ChE-651) seminar by Si
mon Batzner "Equivariant Interatomic Potentials"
DTSTART;VALUE=DATE-TIME:20220524T151500
DTEND;VALUE=DATE-TIME:20220524T161500
UID:4487f228f70a05a77f407298ffd5bd60b0b2ff0afdd7b954ab3e2530
CATEGORIES:Conferences - Seminars
DESCRIPTION:Simon is a PhD student in Applied Mathematics at Harvard. Whil
e interested in far too many things for his own good\, his research focuse
s on building deep learning systems for applications in computational phys
ics and chemistry. Before joining Harvard\, he worked on machine learning
at MIT and on the NASA mission SOFIA. In his free time\, you can find him
playing soccer\, hiking\, and swimming. He comes to Harvard having finishe
d his Master's at MIT. He is originally from beautiful Illertissen\, Germa
ny. \nRepresentations of atomistic structures for machine learning must t
ransform predictably under the geometric transformations of 3D space\, in
particular rotation\, translation\, and reflection as well as permutation
of atoms of the same species. This requirement has traditionally been sati
sfied by operating on descriptors of the atomistic geometry that remain in
variant under the actions of the Euclidean group E(3)\, starting from the
Lennard-Jones potential all the way to today's machine learning approaches
. In this talk\, I will discuss our efforts on generalizing the concept of
invariant representations to the broader class of equivariant representat
ions and will demonstrate how this leads to order-of-magnitude improvement
s in accuracy\, sample efficiency and generalization. The talk will discus
s in particular the Neural Equivariant Interatomic Potential (NequIP) [1]
and show applications to a variety of molecular and materials systems incl
uding small molecules\, water\, a catalytic surface reaction\, glass form
ation of a lithium phosphate\, and Li diffusion in a superionic conductor.
I will discuss then the discovery of a remarkably enhanced power-law be
haviour of the loss of equivariant interatomic potentials. Finally\, I wi
ll show our recent efforts to scale equivariant neural networks to large-
scale systems through the strictly local equivariant interatomic potential
Allegro [2] and show how this allows us to simulate a 100-million-atom Mo
lecular Dynamics simulation. I will conclude by outlining our current theo
retical understanding of equivariant interatomic potentials as well as ope
n research questions. \n\n \n\n[1] In print\, Nature Communications\,
Preprint: arxiv.org/abs/2101.03164\n\n[2] Preprint: arxiv.org/abs/2204.0
5249
LOCATION:https://epfl.zoom.us/j/64473017589?pwd=Vmpnd1pleGhEb1hFb3kxUlNIUW
JyQT09
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR