Barrier Height Prediction by Machine Learning Correction of Semiempirical Calculations.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Accurate prediction of reaction barrier heights is crucial for understanding chemical kinetics.
• Semiempirical quantum mechanical (SQM) methods offer a computationally efficient alternative to high-level methods like density functional theory (DFT).
• Bridging the accuracy gap between SQM and DFT for barrier heights remains a challenge.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting DFT-quality barrier heights (BHs) using SQM calculations.
• To assess the performance of multitask deep neural networks, XGBoost, and Gaussian process regression for this task.
• To identify key features influencing prediction accuracy for future model development.

Main Methods:

• Implementation of three distinct ML models: multitask deep neural network, gradient-boosted trees (XGBoost), and Gaussian process regression.
• Training and validation of models using SQM calculations as input features to predict DFT-quality BHs.
• Analysis of feature importance to identify significant predictors for BHs.

Main Results:

• The developed ML models achieved mean absolute errors comparable to existing methods for the same dataset size.
• Multitask deep neural networks, XGBoost, and Gaussian process regression demonstrated effectiveness in predicting BHs.
• Bespoke predictors were identified as highly impactful features, constituting 70% of the top predictors.

Conclusions:

• The proposed ML corrections offer a viable approach for rapid screening of large reaction networks in combustion and astrochemistry.
• The identified custom-made predictors hold potential for enhancing future Δ-ML models in predicting various reaction properties.
• This work facilitates faster and more accurate computational studies in chemical kinetics and related fields.