Free-energy perturbation in the exchange-correlation space accelerated by machine learning: application to silica...

Area of Science:

• Computational materials science
• Quantum chemistry
• Machine learning in physics

Background:

• Calculating phase transition properties like temperature and entropy is crucial for materials science.
• Existing computational methods often struggle with accuracy for complex systems, particularly those with small transition entropies.
• The 'Jacob's ladder' framework categorizes density functional approximations by increasing accuracy and cost.

Purpose of the Study:

• To develop and validate a machine learning-accelerated free-energy perturbation method for computing transition temperatures and entropies.
• To assess the accuracy of different "rungs" of density functional approximations for the dynamically stabilized phases of SiO2.
• To establish a reliable computational procedure for evaluating and developing new electronic structure functionals.

Main Methods:

• Employed a free-energy-perturbation approach enhanced by machine-learning potentials.
• Applied the method to study the dynamically stabilized phases of silicon dioxide (SiO2).
• Systematically evaluated functionals across rungs 1-4 and the random phase approximation (RPA) on rung 5 of Jacob's ladder.

Main Results:

• Functionals on rungs 1-4 of Jacob's ladder predicted SiO2 transition temperatures with significant errors (25-200%).
• The random phase approximation (rung 5) achieved high accuracy, with a relative error of only 5% for the transition temperature.
• The study demonstrates the necessity of higher-rung methods for accurate predictions in challenging systems.

Conclusions:

• Machine learning-accelerated free-energy perturbation is an efficient approach for calculating transition properties.
• Accurate prediction of transition temperatures in systems like SiO2 requires advanced computational methods, specifically the random phase approximation.
• The study provides a benchmark and methodology for future development and evaluation of electronic structure functionals.