Assessing the Accuracy of Machine Learning Thermodynamic Perturbation Theory: Density Functional Theory and Beyond

Area of Science:

• Computational chemistry
• Materials science
• Machine learning applications

Background:

• Machine learning thermodynamic perturbation theory (MLPT) estimates finite temperature properties by training on limited calculations from a molecular dynamics trajectory.
• MLPT, based on thermodynamic perturbation theory (TPT), can face inaccuracies if the initial trajectory's configuration space poorly overlaps with the target theory's space.

Purpose of the Study:

• To assess the accuracy of MLPT for ensemble total energies and adsorption enthalpies, particularly for molecules in zeolites using various density functional theory (DFT) approximations.
• To identify problematic MLPT cases and demonstrate a method to improve accuracy.

Main Methods:

• Utilized case studies of molecules adsorbed in zeolites with different DFT approximations.
• Employed machine learning-based Monte Carlo (MLMC) resampling to correct for configurational space overlap issues.
• Assessed MLPT accuracy for ensemble total energies and enthalpies of adsorption.

Main Results:

• Identified conditions where MLPT can be inaccurate due to poor configurational space overlap.
• Demonstrated that problematic MLPT cases can be detected without prior knowledge of reference results.
• Showed that MLMC resampling can recover target level results within chemical accuracy even in challenging scenarios.
• Confirmed the accuracy of MLPT for high-level random phase approximation (RPA) calculations of adsorption enthalpies.

Conclusions:

• MLPT is a valuable tool for high-throughput computational chemistry, enabling property prediction at expensive levels of theory.
• MLMC resampling offers a robust strategy to overcome limitations of MLPT related to configurational space sampling.
• The study validates MLPT for accurate prediction of adsorption enthalpies, even at computationally demanding levels like RPA.