We are thrilled to announce the latest publication from our lab, titled "Enhancing Activation Energy Predictions under Data Constraints Using Graph Neural Networks" in the Journal of Chemical Information and Modeling. This study highlights an innovative application of machine learning to predict activation energy, a critical property in chemical reaction engineering.
In this work, we developed a framework using Graph Neural Networks (GNNs) to improve activation energy predictions by incorporating low-cost computational data from semiempirical quantum mechanics (SQM) methods. Traditional methods often require extensive high-level quantum chemistry calculations, which limit their applicability in data-scarce environments. Our approach addresses these challenges by leveraging low-level data to reduce computational costs while maintaining high accuracy.
Key Findings:
-
Enhanced Prediction Accuracy: The delta learning method achieved a significantly lower Mean Absolute Error (MAE) of 3.85 kcal/mol, compared to 6.16 kcal/mol for the baseline GNN model trained exclusively on high-level data.
-
Data Efficiency: Delta learning trained with only 20–30% of high-fidelity data achieved accuracy comparable to methods requiring full datasets, demonstrating its robustness in data-limited scenarios.
-
Broad Applicability: By leveraging low-level data such as GFN2-xTB activation energies, this framework offers a scalable solution for modeling activation energies in diverse chemical systems.
This research underscores the transformative potential of machine learning in chemical engineering, providing a practical and efficient approach for predicting kinetic parameters.
For more details, access the full publication here.