A Euclidean transformer for fast and stable machine learned force fields.

Area of Science:

• Computational chemistry
• Materials science
• Quantum mechanics

Background:

• Machine learned force fields (MLFFs) show promise but face stability issues in long molecular dynamics (MD) simulations.
• Equivariant representations can improve MLFF robustness but are computationally expensive.

Purpose of the Study:

• To develop a computationally efficient and stable MLFF for accurate quantum property analysis.
• To enable reliable, long-timescale MD simulations of large molecular systems.

Main Methods:

• Proposed SO3KRATES, a transformer architecture using sparse equivariant representations and self-attention.
• Separated invariant and equivariant information to avoid costly tensor products.
• Applied SO3KRATES to generate stable MD trajectories and explore potential energy surface (PES) topology.

Main Results:

• SO3KRATES achieves high accuracy, stability, and speed, outperforming existing methods.
• Generated stable MD trajectories for hundreds of atoms in peptides and supramolecular structures.
• Explored thousands of minima for medium-sized molecules, revealing new conformations.

Conclusions:

• SO3KRATES offers a balance between simulation stability and the discovery of novel low-energy conformations.
• Enables insightful analysis of quantum properties for complex systems at unprecedented scales.
• Crucial for realistic molecular exploration tasks in biochemistry and materials science.