AisNet: A Universal Interatomic Potential Neural Network with Encoded Local Environment Features.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Accurate prediction of atomic energies and forces is crucial for materials simulation.
• Existing methods like SchNet have limitations in diverse material systems.

Purpose of the Study:

• To develop a novel interatomic potential energy neural network, AisNet.
• To improve the efficiency and accuracy of predicting atomic properties across various materials.

Main Methods:

• AisNet integrates an autoencoder-based encoding module with embedding, a triplet loss function, and an atomic central symmetry function (ACSF).
• It incorporates an interaction module with periodic boundary conditions (PBC) and a prediction module.
• The model encodes universal local environment features, including elements and atomic positions.

Main Results:

• AisNet demonstrates comparable accuracy to SchNet for molecular datasets (MD17).
• The inclusion of ACSF enhances AisNet's accuracy by 16.8% for energy and 28.6% for force in metal and ceramic datasets.
• AisNet significantly outperforms SchNet and DeepMD in force prediction for alloys, particularly ternary FeCrAl.

Conclusions:

• AisNet efficiently predicts atomic energies and forces across diverse molecular and crystalline materials.
• The model's encoding of local environment features reduces data dependency.
• AisNet shows strong potential for broader applications in materials science with further feature incorporation.