Lightweight and effective tensor sensitivity for atomistic neural networks.

Area of Science:

• Computational chemistry
• Machine learning for materials science

Background:

• Atomistic machine learning models enforce physical symmetries like translation and rotation invariance.
• Current models often rely on scalar invariants (e.g., atom distances).
• There's a growing need for molecular representations using higher-rank rotational tensors.

Purpose of the Study:

• To introduce a framework extending Hierarchically Interacting Particle Neural Network (HIP-NN) with Tensor Sensitivity (TS) information.
• To enable direct incorporation of many-body information with minimal parameter increase.
• To evaluate the accuracy and performance improvements of the new HIP-NN-TS model.

Main Methods:

• Developed HIP-NN-TS by integrating Tensor Sensitivity information into local atomic environments.
• Employed a weight tying strategy for efficient inclusion of many-body interactions.
• Tested HIP-NN-TS on various datasets and network sizes, comparing against HIP-NN and other models.

Main Results:

• HIP-NN-TS demonstrates superior accuracy compared to HIP-NN across multiple datasets and network sizes.
• The accuracy improvement from tensor sensitivities becomes more pronounced with complex datasets.
• Achieved a record mean absolute error of 0.927 kcal/mol for conformational energy on the COMP6 benchmark.

Conclusions:

• HIP-NN-TS offers a significant accuracy enhancement over HIP-NN with a negligible increase in model parameters.
• The framework effectively leverages tensor sensitivities for improved molecular property prediction.
• HIP-NN-TS represents a state-of-the-art approach for atomistic machine learning, particularly for complex molecular systems.