Accurately estimating activation energies by leveraging neural network methods and a large dataset

Determining activation energies is integral to the field of computational chemistry. With the emergence of artificial intelligence, new methodologies such as neural networks have been introduced to accelerate the prediction of these energies, representing a notable advancement in this scientific domain. By incorporating topological indices, molecular fingerprints of reactants and products, and reaction enthalpy as descriptors, a deep-learning framework was developed. This framework utilizes the Reaction Graph Depth 1 (RGD1) dataset, which includes 176,992 organic reactions, to accurately estimate activation energies using artificial neural networks. The results demonstrated training R² values of 0.99, with a mean absolute error (MAE) of 2.06 kcal/mol and a root mean square error (RMSE) of 3.20 kcal/mol across an activation energy range of nearly 200 kcal/mol. These results exceed the accuracy of the other models on the same dataset as well as different datasets. Based on the learning curve, the training and validation losses were nearly identical and minimized, suggesting that the model was effectively regularized. The Chemprop model, with optimized hyperparameters, reached an R² of 0.93 on the test set, which is slightly below the performance of the previously discussed ANN method.