Rapidly predicting Kohn-Sham total energy using data-centric AI.

Area of Science:

• Computational Materials Science
• Materials Informatics
• Quantum Chemistry

Background:

• Density Functional Theory (DFT) and DFT-based Tight Binding (DFTB) calculations are crucial for predicting material properties.
• However, these methods face limitations in computational cost, especially for large systems and varying temperatures.
• Accurate prediction of electronic structures is essential for understanding material behavior.

Purpose of the Study:

• To introduce a novel, data-centric machine learning framework for rapid and accurate prediction of Kohn-Sham (KS) total energy.
• To demonstrate the framework's ability to predict properties of anatase TiO2 nanoparticles (NPs) at different temperatures.
• To eliminate the need for experimental data in materials property prediction.

Main Methods:

• A data-centric machine learning framework, termed co-modeling, was developed.
• The framework utilizes a small amount of theoretical data to train the model.
• It is designed to be general and applicable to various nanoparticles (NPs).

Main Results:

• The co-modeling framework accurately predicts the KS total energy of anatase TiO2 NPs.
• The approach significantly reduces the computational run-time compared to traditional DFT/DFTB methods.
• The framework demonstrates effectiveness across different temperatures without requiring experimental data.

Conclusions:

• The novel co-modeling framework offers a computationally efficient alternative for predicting material properties.
• This approach accelerates the study of electronic structures and related physical/chemical properties of nanoparticles.
• A web service was developed to showcase the framework's practical application and effectiveness.