Many-Body Function Corrected Neural Network with Atomic Attention (MBNN-att) for Molecular Property Prediction

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Machine learning (ML) models are increasingly used for chemical property prediction.
• Existing models often lack general-purpose, high-performance capabilities accessible on central processing unit (CPU) devices.
• A need exists for low-cost, efficient ML models for broad chemical applications.

Purpose of the Study:

• To introduce a novel ML model, MBNN-att, for predicting molecular and material properties.
• To achieve high performance and generalizability on CPU devices.
• To enhance the accuracy and transferability of ML models in chemistry.

Main Methods:

• Developed the MBNN-att model by incorporating an atomic attention mechanism into a many-body neural network (MBNN).
• Utilized explicit function descriptors as inputs for an atom-based feed-forward neural network (NN).
• Implemented a multihead self-attention mechanism using a vector output from the NN, splitting it into atomic attention weights and many-body functions.

Main Results:

• MBNN-att demonstrated excellent performance on all QM9 properties, achieving errors below chemical accuracy.
• The model particularly excelled in predicting energy-related extensive properties.
• Systematic comparisons confirmed the superiority of MBNN-att over other descriptor-based and graph-representation ML models.

Conclusions:

• The many-body function framework and atomic attention mechanism are crucial for the high performance of MBNN-att.
• MBNN-att exhibits strong transferability in molecular property prediction.
• The developed model offers a promising solution for accessible, high-performance ML in computational chemistry.