Modelling local and general quantum mechanical properties with attention-based pooling.

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Mechanics

Background:

• Atom-centered neural networks (ACNNs) are state-of-the-art for approximating molecular quantum chemical properties.
• Current ACNNs often use simple sum or average pooling, which may limit representational power for localized or intensive properties.
• Existing pooling methods may not fully capture complex interatomic interactions crucial for accurate predictions.

Purpose of the Study:

• To introduce a learnable, attention-based pooling mechanism for ACNNs.
• To enhance the conversion from atomic to molecular representations in deep learning models for chemistry.
• To improve the prediction accuracy of molecular properties by better modeling atom interactions.

Main Methods:

• Developed a novel attention-based pooling operation as a drop-in replacement for existing methods.
• Integrated the attention pooling into established ACNN architectures like SchNet and DimeNet++.
• Evaluated performance on diverse datasets, molecular properties, and levels of theory.

Main Results:

• The proposed attention pooling consistently outperformed sum, mean, and a physics-aware pooling method.
• Achieved significant performance uplifts, with improvements up to 85% on specific tasks.
• Demonstrated the effectiveness of learnable pooling in capturing complex atomic interactions.

Conclusions:

• Attention-based pooling offers a superior alternative to traditional methods for molecular property prediction.
• This approach enhances the representational capacity of ACNNs without altering core architectural components.
• The developed pooling mechanism represents a significant advancement in geometric deep learning for quantum chemistry.