SchNet_IIA: Potential Energy Surface Fitting by Interatomic Interactions Attention Based on Transfer Learning...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning (ML) accelerates molecular dynamics (MD) simulations and potential energy surface (PES) fitting.
• Existing ML models, like artificial neural networks (ANNs) and multilayer perceptrons (MLPs), use basic architectures and lack interpretability.
• Advanced ML techniques are needed to overcome limitations in current models.

Purpose of the Study:

• To develop a novel model analysis method for direct causal analysis of ML models.
• To enhance the interpretability and performance of ML frameworks for PES fitting and MD simulations.
• To propose a generalized approach for ML model analysis in computational science.

Main Methods:

• Developed a feature-representation-transfer approach for causal analysis of ML models.
• Analyzed the SchNet framework by constructing diverse source tasks.
• Introduced interatomic interactions attention (IIA) for characterizing doped clusters.

Main Results:

• Achieved a 0.015 eV/atom accuracy enhancement over the original SchNet model.
• Significantly improved the model's ability to capture atomic environment characteristics.
• Smoothed activation functions led to a 23.47% increase in convergence speed.
• The SchNet_IIA model demonstrated superior performance in capturing interatomic interactions.

Conclusions:

• The proposed transfer learning analysis method offers a new perspective for ML model interpretability.
• The SchNet_IIA model shows enhanced accuracy and efficiency in molecular simulations.
• This work presents a valuable, generalizable approach for analyzing complex ML models in scientific applications.