Learning Rates: Arrhenius Modeling for Hydrogen Abstraction Reactions

This repository contains the implementation accompanying the article:

Learning Rates: Predicting Reaction Rate Coefficients for Hydrogen Abstraction Reactions

It provides a Chemprop-based reaction modeling framework extended with an Arrhenius-parameter prediction workflow tailored for hydrogen abstraction kinetics.

Overview

The core objective of this project is to predict full Arrhenius parameter triplets

$k(T) = A T^n \exp!\left(-\frac{E_a}{RT}\right)$

for forward and reverse hydrogen abstraction reactions using graph neural networks.

This repository introduces:

• Arrhenius-specific model heads and loss formulations
• Reaction-aware data preprocessing and role labeling
• Donor-acceptor ordered and invariant training modes
• Forward/reverse thermodynamic consistency handling
• Temperature-grid evaluation of log10 k(T)
• Custom cross-validation and splitting strategies for reaction kinetics

All Arrhenius-specific logic is isolated in:

arrhenius/

The upstream Chemprop framework is used as the message-passing backbone and training infrastructure.

Repository Structure

arrhenius/
    preprocessing/      # Reaction-specific data preparation
    training/           # Training loops and HPO workflow
    models/             # Arrhenius heads and layers
    evaluation/         # Temperature-grid evaluation
    README.md           # Entry point for Arrhenius workflow

If you are interested specifically in the kinetics workflow, begin with:

• arrhenius/README.md
• arrhenius/preprocessing/README.md
• arrhenius/training/hpo/WORKFLOW.md

Relationship to Chemprop

This project builds on Chemprop v2, which provides:

• Directed message passing neural network backbone
• Molecular/reaction featurization
• Training infrastructure and CLI
• Lightning-based training utilities

The Arrhenius extension adds:

• Parameterized Arrhenius output heads
• Kinetics-specific objective functions
• Reaction-level splitters
• Hydrogen abstraction-specific preprocessing
• Thermodynamic coupling of forward/reverse directions

The Chemprop core remains structurally recognizable. All domain-specific logic is confined to arrhenius/.

Chemprop documentation: https://chemprop.readthedocs.io/en/main/

Scientific Scope

This repository supports:

• Prediction of forward and reverse Arrhenius parameters
• Evaluation of log10 k(T) over 300-3000 K
• Analysis by donor-acceptor reactive site types
• Uncertainty-aware modeling (when enabled)
• Cross-validation protocols appropriate for reaction kinetics

The workflow is designed specifically for hydrogen abstraction reactions and is not a general-purpose property prediction package.

Installation

Installation follows standard Chemprop v2 setup procedures.

After installation, the Arrhenius workflow can be accessed via the modules under arrhenius/.

Citation

If this repository contributes to your work, please cite:

Learning Rates: Predicting Reaction Rate Coefficients for Hydrogen Abstraction Reactions

and the relevant Chemprop references:

Yang et al., Chemprop v2: An Efficient, Modular Machine Learning Package for Chemical Property Prediction, J. Chem. Inf. Model. (2025).

Yang et al., Chemprop: A Machine Learning Package for Chemical Property Prediction, J. Chem. Inf. Model. (2023).

License

Chemprop is distributed under the MIT License. Arrhenius-specific extensions follow the same licensing structure.