Enhancing Activation Energy Predictions under Data Constraints Using Graph Neural Networks.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemical Engineering

Background:

• Accurate activation energy prediction is vital for chemical reaction modeling.
• High computational costs of quantum chemistry limit data availability.
• Machine learning offers a potential solution to data scarcity.

Purpose of the Study:

• To compare transfer learning, delta learning, and feature engineering for enhancing activation energy predictions.
• To evaluate these methods using low-cost, semiempirical quantum mechanics (SQM) data with graph neural networks (Chemprop).
• To identify strategies balancing prediction accuracy and computational cost.

Main Methods:

• Systematic evaluation of three machine learning approaches: transfer learning, delta learning, and feature engineering.
• Utilizing graph neural networks (Chemprop) with data from semiempirical quantum mechanics (SQM) calculations.
• Comparing performance against high-level CCSD(T)-F12a targets.

Main Results:

• Delta learning proved most effective, achieving high accuracy with significantly less data (20-30%) compared to full datasets.
• Transfer learning showed variable results, sensitive to reaction distribution mismatches.
• Feature engineering provided modest improvements, particularly for thermodynamic properties.

Conclusions:

• Delta learning offers a powerful, data-efficient method for activation energy prediction, despite computational demands in application.
• The choice of method involves a trade-off between accuracy and computational resources.
• Findings provide guidelines for applying machine learning in chemical reaction engineering under resource constraints.