Learning local equivariant representations for large-scale atomistic dynamics.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning for science

Background:

• Accurate and efficient potential energy surface (PES) parametrization is crucial for molecular and materials simulations.
• Atom-centered message passing neural networks (MPNNs) offer high accuracy but are limited in accessible length-scales.
• Traditional local methods scale well but lack accuracy.

Purpose of the Study:

• Introduce Allegro, a novel deep neural network architecture for interatomic potentials.
• Develop a strictly local equivariant model that balances accuracy and computational efficiency.
• Demonstrate Allegro's capability for large-scale molecular simulations.

Main Methods:

• Allegro utilizes iterated tensor products of learned equivariant representations, avoiding atom-centered message passing.
• The architecture is a strictly local equivariant deep neural network.
• Performance is evaluated on QM9 and revMD17 datasets and through molecular dynamics simulations.

Main Results:

• Allegro achieves superior accuracy and scalability compared to state-of-the-art methods on benchmark datasets.
• A single tensor product layer in Allegro surpasses existing deep MPNNs and transformers on QM9.
• Allegro demonstrates excellent generalization to out-of-distribution data and accurately predicts properties of amorphous electrolytes.

Conclusions:

• Allegro offers a breakthrough in developing accurate and computationally efficient interatomic potentials.
• The model's local equivariant nature enables high-fidelity simulations at unprecedented scales.
• Allegro paves the way for advanced molecular simulations in diverse scientific domains.