Neural network-based pseudopotential: development of a transferable local pseudopotential.

Area of Science:

• Computational Materials Science
• Quantum Chemistry
• Machine Learning Applications

Background:

• Fast quantum simulations of materials rely on transferable local pseudopotentials (LPPs).
• Existing LPPs often lack transferability, particularly those not adhering to norm-conserving conditions.
• This limitation hinders accurate predictions across different chemical environments.

Purpose of the Study:

• To develop a novel deep neural network (DNN) approach for generating highly transferable LPPs.
• To address the limitations of conventional LPPs by incorporating norm-conserving conditions.
• To enhance the accuracy and predictive power of pseudopotential calculations.

Main Methods:

• Introduced a generalized Kerker method integrated with a DNN to represent norm-conserving pseudowavefunctions.
• Formulated pseudopotential conditions within a loss function for DNN training.
• Utilized back-propagation with single-point all-electron atom data for minimization.

Main Results:

• Generated neural network-based LPPs (NNLPs) for s- and p-block elements (periods 2-4).
• NNLPs demonstrated superior transferability and accuracy compared to conventional LPPs in atomic calculations.
• NNLPs showed enhanced predictive accuracy for bulk properties, including binary alloys, surpassing specialized bulk LPPs.

Conclusions:

• The DNN-based approach successfully produces highly transferable LPPs (NNLPs).
• NNLPs offer a significant advancement for accurate and efficient quantum simulations of materials.
• This method provides a robust framework for developing improved pseudopotentials applicable across diverse systems.