TrIP2: Expanding the Transformer Interatomic Potential Demonstrates Architectural Scalability for Organic Compounds

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning for Science

Background:

• Transformer interatomic potentials (TrIP) offer promising transferable models for molecular simulations.
• Existing models may require significant adaptation for new chemical elements or molecular configurations.

Purpose of the Study:

• To introduce TrIP2, an enhanced interatomic potential based on the TrIP architecture.
• To evaluate TrIP2's accuracy, transferability, and performance across diverse chemical systems and tasks.

Main Methods:

• TrIP2 utilizes an equivariant SE(3)-transformer architecture trained on an expanded dataset (ANI-2x) including sulfur, fluorine, and chlorine.
• Performance was benchmarked on energy/force calculations (COMP6), structure minimization, and torsion drives.
• Direct comparisons were made against ANI-2x, AIMNet2, and MACE-OFF23.

Main Results:

• TrIP2 demonstrates high accuracy and transferability to new atom types without architectural changes.
• Achieved state-of-the-art force prediction on COMP6 benchmarks.
• Approached DFT-optimized structures in geometry optimization and torsion drive tasks.

Conclusions:

• TrIP2 offers enhanced generalizability and precision by leveraging expanded training data.
• The model establishes a robust, scalable framework for future applications and domain expansions with minimal reengineering.