Informing geometric deep learning with electronic interactions to accelerate quantum chemistry.

Area of Science:

• Computational Chemistry
• Materials Science
• Drug Discovery

Background:

• Predicting chemical properties is crucial for discovering new materials and medicines.
• Machine learning methods struggle with limited training data in unexplored chemical spaces.
• Existing computational methods like density functional theory are computationally expensive.

Purpose of the Study:

• To develop a novel deep learning approach for accurate prediction of molecular electronic energies and properties.
• To overcome the challenge of data scarcity in machine learning for chemical applications.
• To create a method that is significantly faster than traditional computational chemistry techniques.

Main Methods:

• Developed a physics-inspired equivariant neural network, OrbNet-Equi.
• Incorporated molecular electronic structure and electronic interactions among atomic orbitals into deep learning.
• Utilized efficient tight-binding simulations and learned mappings to predict physical quantities.

Main Results:

• OrbNet-Equi accurately models a wide spectrum of chemical properties.
• The method is orders of magnitude faster than density functional theory.
• Outperformed traditional semiempirical and machine learning methods on diverse chemical processes, including charge-transfer complexes and open-shell systems.

Conclusions:

• OrbNet-Equi effectively addresses data scarcity by integrating physical knowledge into deep learning.
• The method offers a computationally efficient and accurate alternative for predicting chemical properties.
• This approach expands possibilities for chemistry and materials science research by reducing the cost of data acquisition.