Uncertainty-Aware Machine Learning Models of Many-Body Atomic Interactions

Explore a comprehensive lecture on uncertainty-aware machine learning models for many-body atomic interactions presented by Boris Kozinsky from Harvard University. Delve into the evolution of atomistic modeling methods, limitations of current ML functionals, and innovative approaches like nonlocal features for exchange energy. Discover symmetry-aware machine learning force fields, including E(3)-equivariant neural interatomic potentials and the Allegro architecture. Examine FLARE Bayesian Force Fields and their application in on-the-fly active learning. Investigate the use of ML force fields for transition metals and micron-scale heterogeneous reaction dynamics. Gain insights into selecting optimal training sets and simulating large-scale dynamics, including phase transitions in 2D stanene and the evolution of Li anode-electrolyte interfaces.

Intro
Machine learning for understanding dynamics
Atomistic modeling methods evolution
Limitations of Current ML Functionals
Nonlocal Features for Exchange Energy
Model and Training Details
What is the derivative discontinuity?
Dynamic problems require dynamic solutions
Computing forces for molecular dynamics
Symmetry-Aware Machine Learning Force Fields
E(3) equivariance allows to capture 3D geometry
NequIP: E(3)-equivariant Neural Interatomic Potentials
Long MD simulation stability
Allegro's two-track architecture
Allegro accuracy and scalability
Allegro: Large-scale dynamics
Selecting optimal training sets
FLARE Bayesian Force Fields
FLARE on the fly active learning
Phase transitions in 2D stanene
ML force fields for transition metals
Micron-scale heterogeneous reaction dynamics
Evolution of Li anode-electrolyte interface