E(n)-Equivariant cartesian tensor message passing interatomic potential.

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Quantum Chemistry

Background:

• Machine learning potentials (MLPs) are increasingly used to approximate computationally expensive first-principles calculations for large systems.
• Message passing neural networks (MPNNs) show high accuracy, with many recent models based on Cartesian coordinates.
• Existing MPNNs often limit node information to scalars and vectors, restricting their representational capacity.

Purpose of the Study:

• To introduce High-order Tensor message Passing interatomic Potential (HotPP), an E(n) equivariant MPNN.
• To extend node embeddings and messages to arbitrary-order tensors for enhanced information representation.
• To enable direct prediction of high-order tensor properties like dipole moments and polarizabilities.

Main Methods:

• Developed an E(n) equivariant message passing neural network architecture, HotPP.
• Incorporated arbitrary-order tensors for node embeddings and messages.
• Utilized basic equivariant operations to couple high-order tensors.
• Applied HotPP to predict target properties and calculate various spectra.

Main Results:

• HotPP achieves high accuracy in predicting target properties across multiple datasets.
• The model successfully calculates phonon spectra, infrared spectra, and Raman spectra.
• HotPP demonstrates the ability to directly predict high-order tensors without architectural modifications.

Conclusions:

• HotPP offers a significant advancement in MLPs by leveraging high-order tensor representations.
• The model's versatility extends to predicting molecular properties and spectroscopic data.
• HotPP shows great potential as a versatile tool for future research in computational materials science and chemistry.