Abstract
Fast and accurate estimation of reaction activation energies is crucial for understanding chemical reactivity and accelerating catalyst discovery. While traditional methods based on quantum chemical calculations are reliable, they are computationally expensive, motivating the development of machine learning approaches. In this study, we propose lightweight frameworks based on Crystal Graph Convolutional Neural Networks (CGCNN) with about a tenth of the number of parameters of state-of-the-art graph neural network models to predict activation energies for both organic molecular reactions and heterogeneous surface reactions. Two model architectures̶a Kronecker product-based model and a difference-based model - were investigated using the Transition1x and Open Catalyst Project (OCP) datasets. The Kronecker product model consistently outperformed the difference-based model, achieving a coefficient of determination (𝑅2) of 0.45 on the Transition1x dataset which is generally better than those of methods (𝑅2 ∼ 0.3) based on a molecular descriptor and a decision tree model. We also investigated the Open Catalyst Project (OCP) dataset to identify challenges for estimating activation energies in heterogeneous catalytic systems.