BMach: a Bayesian machine for optimizing Hubbard U parameters in DFT+U with machine learning

Ritwik Das

2024-10-23

Abstract:Accurately determining the effective Hubbard parameter $(U_{eff})$ in Density Functional Theory plus U (DFT+U) remains a significant challenge, often relying on empirical methods or linear response theory, which frequently fail to predict accurate material properties. This study introduces BMach, an advanced Bayesian optimization algorithm that refines $U_{eff}$ by incorporating electronic properties, such as band gaps and eigenvalues, alongside structural properties like lattice parameters. Implemented within the Quantum Espresso platform, BMach demonstrates superior accuracy and reduced computational cost compared to traditional methods. The BMach-optimized $U_{eff}$ values yield electronic properties that align closely with experimental and high-level theoretical results, providing a robust framework for high-throughput materials discovery and detailed electronic property characterization across diverse material systems.

Computational Physics,Materials Science,Strongly Correlated Electrons

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve

The paper aims to address the challenge of accurately determining the effective Hubbard parameter (Ueff) in Density Functional Theory plus U (DFT+U). Traditional empirical methods or Linear Response Theory (LR) often fail to accurately predict material properties, especially when dealing with materials with strong electron correlation effects. These issues include:

1. **Limitations of Empirical Methods**: Empirical methods rely on experimental data, but when such data is unavailable, these methods become ineffective.
2. **High Computational Cost of Linear Response Theory**: While Linear Response Theory can provide relatively accurate results, its computational cost is high, particularly when dealing with large-scale systems.

To address these issues, the study introduces a Bayesian optimization algorithm named BMach. BMach optimizes Ueff by combining electronic properties (such as band gap and eigenvalues) and structural properties (such as lattice parameters). Specifically, the main contributions of BMach include:

- **Improved Accuracy**: The Ueff values optimized by BMach can make electronic properties highly consistent with experimental and advanced theoretical results.
- **Reduced Computational Cost**: Compared to traditional methods, BMach can explore the parameter space more efficiently and quickly converge to the optimal solution.
- **Multi-objective Optimization**: BMach can simultaneously consider multiple objectives such as band gap, eigenvalues, and lattice parameters, providing a more comprehensive prediction of material properties.

Through these improvements, BMach provides a powerful framework for high-throughput material discovery and detailed electronic property characterization.