Transferable and Transparent Energy Decomposition-Based Machine Learning Models for Computing Accurate Reaction...

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Chemistry

Background:

• Density Functional Theory (DFT) is widely used for chemical reaction energy calculations.
• Standard DFT methods can suffer from accuracy limitations.
• Improving the precision of these calculations is crucial for predicting chemical behavior.

Purpose of the Study:

• To develop a machine-learning framework to enhance DFT reaction energy accuracy.
• To create a transferable, interpretable, and modular approach.
• To provide a robust alternative to standard DFT and complex neural networks.

Main Methods:

• Decomposing DFT reaction energies into physically meaningful, chemically intuitive contributions.
• Training linear regression (LR) models using these decomposed energy descriptors.
• Employing a random forest (RF) classifier to dynamically select the optimal LR model.
• Utilizing physically meaningful energy-decomposition descriptors.

Main Results:

• A general-purpose LR model reduced mean absolute percentage errors (MAPE) by up to 63% compared to uncorrected DFT.
• Specialized LR models further improved accuracy for specific reaction types.
• The RF/LR pipeline achieved a MAPE reduction of up to 123 percentage points.
• The framework demonstrated robust performance on out-of-distribution data, including transition-metal complexes.

Conclusions:

• The developed machine-learning framework significantly enhances DFT reaction energy accuracy.
• The approach is transferable, interpretable, and maintains performance on unseen data.
• This method offers a reliable and interpretable alternative for accurate chemical reaction energy predictions.