Learning molecular energies using localized graph kernels.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning models can now predict atomic potential energy with high accuracy, suitable for molecular dynamics.
• Encoding physical constraints like symmetry into these models is crucial for optimal performance.
• Existing methods struggle to efficiently incorporate translation, rotation, and permutation symmetries.

Purpose of the Study:

• To develop a novel machine learning approach that inherently integrates physical symmetries for energy modeling.
• To present a graph theory-based method for representing atomic environments and their similarities.
• To benchmark the proposed method's accuracy in predicting molecular atomization energies.

Main Methods:

• Utilized graph theory, specifically a random walk graph kernel, to compare local atomic environments represented by adjacency matrices.
• Developed the Graph Approximated Energy (GRAPE) approach, designed to naturally handle translation, rotation, and permutation symmetries.
• Applied a simple version of GRAPE to predict atomization energies for organic molecules.

Main Results:

• The GRAPE approach successfully incorporates essential physical symmetries into machine learning models.
• The method demonstrates flexibility and potential for various extensions.
• Benchmarking showed accurate prediction of atomization energies on a standard dataset.

Conclusions:

• The graph theory-based GRAPE method offers an effective way to encode physical symmetries in machine learning for energy modeling.
• This approach has significant implications for accelerating molecular dynamics simulations and materials discovery.
• GRAPE provides a robust and extensible framework for developing accurate and efficient atomic energy models.