Learning Local Equivariant Representations for Large-Scale Atomistic Dynamics

Explore a comprehensive lecture on Learning Local Equivariant Representations for Large-Scale Atomistic Dynamics presented by Albert Musaelian from Valence Labs. Delve into the world of machine learning interatomic potentials, focusing on the trade-offs between accuracy and speed. Discover the Allegro architecture, a strictly local equivariant deep learning interatomic potential designed for parallel scalability and increased computational efficiency. Examine its applications and benchmarks on various materials and chemical systems, including large-scale biomolecular simulations. Learn about E(3)-equivariant architectures, message passing networks, and the importance of locality in atomistic modeling. Gain insights into scaling techniques for large systems and participate in a Q&A session to further understand this cutting-edge approach to molecular dynamics simulations.

- Intro
- Machine Learning Potentials
- No Constraints, Invariance, Equivariance
- NequIP Generalizes Across Geometry
- Message Passing Networks
- Allegro: Strictly Local Deep Equivariant Model
- Importance of Locality
- Demonstrating Allegro Scaling Up
- Weak Scaling
- Q+A