Bayesian optimization for tuning and selecting hybrid-density functionals

Author:
  • RVRodrigo A. Vargas-Hernández
    Rodrigo A. Vargas-Hernández

Abstract
Hybrid-density functional methods such as B3LYP are widely used across different fields to predict physical properties of molecular and material systems. The optimization of hybrid-density functionals is computationally demanding. The machine learning algorithm presented here, Bayesian optimization, is capable of optimizing efficiently the exchange-correlation constants for hybrid-density functionals and select the exchange-correlation functional that describes a benchmark data most accurately. Bayesian optimization is an iterative method that uses Gaussian process models to find the minimum of non-analytic functions. The same numerical procedure can be applied to other computational methods whose parameters can be tuned by comparison with benchmark data.