Accurately Predicting Barrier Heights for Radical Reactions in Solution Using Deep Graph Networks.

Area of Science:

• Computational Chemistry
• Chemical Kinetics
• Machine Learning in Chemistry

Background:

• Accurate prediction of reaction barriers and solvent effects is crucial for developing kinetic mechanisms and predicting chemical reaction outcomes.
• Existing methods often require computationally intensive optimizations for predicting reaction energetics.

Purpose of the Study:

• To create a comprehensive dataset of elementary radical reactions with associated energetic data.
• To develop a machine learning model capable of directly predicting Gibbs free energies of activation and reaction in both gas and solution phases.
• To enable rapid, high-throughput prediction of reaction properties for applications in synthesis planning.

Main Methods:

• Generated a dataset of 5,600 unique elementary radical reactions using high-level quantum chemical calculations.
• Calculated Gibbs free energies of activation and reaction in 40 solvents using COSMO-RS for solvation.
• Trained a deep graph network model using atom-mapped SMILES strings of reactants, products, and solvents as input.

Main Results:

• The deep graph network achieved a mean absolute error of 1.16 kcal mol⁻¹ for predicting Gibbs free energy of activation in solution.
• The model successfully predicts reaction energetics using only SMILES representations, bypassing the need for structural optimizations during inference.
• The dataset, augmented to ~2 million entries, includes atom-mapped SMILES and verified transition states.

Conclusions:

• The developed deep graph network model offers a computationally efficient approach for predicting radical reaction energetics in solution.
• This work provides a valuable benchmark dataset for future studies in computational chemistry and machine learning.
• The model is well-suited for accelerating high-throughput screening and aiding in (retro-)synthesis planning.