Machine Learning Adaptive Basis Sets for Efficient Large Scale Density Functional Theory Simulation.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Traditional atom-centered basis sets lack adaptability to local atomic environments.
• Adaptive basis sets offer higher accuracy and efficiency but are computationally demanding to generate.

Purpose of the Study:

• To develop a machine learning approach for predicting adaptive basis sets using only local geometric information.
• To demonstrate the efficiency and accuracy of machine learning-predicted adaptive basis sets in density functional theory (DFT) calculations.

Main Methods:

• Utilizing machine learning to predict rotation matrices that transform standard atom-centered basis sets into adaptive ones.
• Employing a rotationally invariant parametrization based on potentials anchored on neighboring atoms.
• Applying the method to molecular dynamics (MD) simulations of liquid water.

Main Results:

• Machine learning successfully predicts adaptive basis sets from local geometry.
• Adaptive basis sets significantly reduce computational cost (up to 200x) and operations (4 orders of magnitude) in DFT.
• Minimal basis sets yield structural properties in good agreement with converged results.

Conclusions:

• Machine learning provides an efficient route to generating accurate adaptive basis sets.
• This method drastically lowers computational expense for DFT, enabling large-scale simulations.
• The approach is robust, even with small training datasets, due to its variational nature.