Efficient Corrections for DFT Noncovalent Interactions Based on Ensemble Learning Models.

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Chemistry

Background:

• Machine learning models are typically database-dependent, requiring retraining for new datasets.
• Accurate computation of noncovalent interactions (NCIs) using first-principles methods is computationally intensive.
• A universal model is needed to handle diverse chemical databases efficiently.

Purpose of the Study:

• To develop a general procedure for constructing ensemble learning models for noncovalent interaction (NCI) databases.
• To evaluate the performance of different ensemble learning frameworks (Bagging, Boosting, Stacking) for NCI prediction.
• To achieve high accuracy in NCI computation with reduced computational cost.

Main Methods:

• Exploration of Bagging, Boosting, and Stacking ensemble learning frameworks.
• Utilizing low-level density functional theory (DFT) calculations on benchmark NCI databases (S66, S22, X40).
• Comparison of ensemble models against single machine learning models and high-level quantum chemical benchmarks (CCSD(T)/CBS).

Main Results:

• Ensemble learning models significantly improved NCI accuracy, achieving a root-mean-square error (RMSE) of 0.22 kcal/mol.
• Ensemble models demonstrated superior accuracy (RMSE lowered by ~25%), robustness, and goodness-of-fit compared to single models.
• Heterogeneous Stacking ensemble models showed the highest application potential.

Conclusions:

• A standardized procedure for constructing ensemble learning models was successfully applied to NCI datasets.
• Ensemble learning provides an efficient and accurate solution for computational chemistry problems, particularly NCI prediction.
• The developed procedure is potentially applicable to other chemical databases beyond NCIs.