TrIP─Transformer Interatomic Potential Predicts Realistic Energy Surface Using Physical Bias.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate interatomic energies and forces are crucial for molecular simulations.
• Machine learning models offer rapid and accurate estimations of these properties.
• Developing universal potentials for diverse atomic species remains a challenge.

Purpose of the Study:

• To introduce the Transformer Interatomic Potential (TrIP), a novel machine learning model for interatomic interactions.
• To demonstrate TrIP's capability in achieving high accuracy for energy and force predictions.
• To advance the development of universal interatomic potentials.

Main Methods:

• Utilized the SE(3)-Transformer architecture for a chemically sound potential.
• Employed a species-agnostic design with continuous atomic representation and graph convolutions.
• Incorporated physical biases, including Ziegler-Biersack-Littmark screening and constrained atomization energies.

Main Results:

• Achieved state-of-the-art accuracy on the COMP6 benchmark with 1.02 kcal/mol MAE for energy prediction.
• Demonstrated improved long- and near-range interactions in water molecule simulations.
• Showcased stability in molecular dynamics simulations, including exploration of Ramachandran space.

Conclusions:

• TrIP represents a significant step towards accurate universal interatomic potentials.
• The species-agnostic architecture facilitates parameter sharing and generalization.
• TrIP offers a promising alternative to existing neural network potentials for molecular simulations.