Unified Understanding of E(3)-Equivariant Interatomic Potentials - Theory and Applications

Explore a comprehensive lecture on E(3)-Equivariant Interatomic Potentials theory and applications presented by Ilyes Batatia from the University of Cambridge. Delve into the world of interatomic potentials, symmetries, and machine learning approaches in quantum mechanics. Discover the unified understanding of ML potentials, including Message Passing Neural Networks, body-ordered messages, and the Generalized Atomic Cluster Expansion (ACE). Examine the Multi-ACE framework, MPNN-ACE identification, and the classification of models within this unified design space. Investigate the BOTNet as a body-ordered Equivariant MPNN and explore the influence of non-linearities and data normalization. Address ML Interatomic Potentials limitations and learn how many-body messages can solve MPNN limitations. Gain insights into data efficiency, extrapolation capabilities, and practical applications such as H transfer in acetyl-acetone. Conclude with an outlook on future developments in this field.

Interatomic potentials (force fields)
Symmetries in Interatomic Potentials
Machine Learning Interatomic Potentials
A Unified Understanding of ML Potentials
Message Passing Neural Networks
Body ordered messages
Generalized Atomic Cluster Expansion (ACE)
Multi-ACE: A Framework of Many-Body Equivariant MPNNS
MPNNS - ACE identification
Classifying models in the Multi-ACE framework
MPNNs as a sparsification of local models
Understanding Nequl in the unified design space
BOTNet : A body ordered Equivariant MPNN
Influence of non-linearities
Data Normalization
ML Interatomic Potentials limitations
Solving MPNNs limitations with many body messages
Required number of message passing
Higher order messages change the learning law
High accuracy on benchmarks
Data Efficiency
Extrapolation and speed
Acetyl-acetone: H transfer
Outlook