CoeffNet: predicting activation barriers through a chemically-interpretable, equivariant and physically constrained...

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Chemical Kinetics

Background:

• Calculating activation barriers is crucial for understanding reaction mechanisms and kinetics.
• Traditional electronic structure methods for computing activation barriers are computationally intensive and time-consuming.
• Identifying transition states is a bottleneck in predicting reaction rates.

Purpose of the Study:

• Introduce CoeffNet, an equivariant graph neural network for predicting activation barriers.
• Utilize frontier molecular orbital coefficients as graph node features for enhanced interpretability and physical constraints.
• Demonstrate the model's capability on S_N2 reactions as a proof-of-concept.

Main Methods:

• Developed CoeffNet, an equivariant graph neural network architecture.
• Employed coefficients of frontier molecular orbitals (e.g., highest occupied molecular orbital) from reactant and product complexes as input features.
• Trained and validated the model on a dataset of S_N2 reactions.

Main Results:

• CoeffNet accurately predicts activation barriers with a mean absolute error below 0.025 eV.
• The model provides chemically interpretable outputs, including transition state molecular orbital coefficients.
• Visualization of highest occupied molecular orbital densities in transition states offers insights into reaction pathways.

Conclusions:

• CoeffNet offers a computationally efficient and interpretable alternative for predicting activation barriers.
• The use of molecular orbital coefficients as features enhances the physical relevance and chemical intuition of the model's predictions.
• This approach shows significant promise for accelerating the study of molecular reaction mechanisms and kinetics.