Predicting Enthalpy of Formation of Energetic Compounds by Machine Learning: Comparison of Featurization Methods and Algorithms

Machine learning (ML) is an emerging approach for predicting molecular properties. The prediction of the properties of energetic molecules by ML is still in its infancy. In order to improve the accuracy of ML-based predictions, it is important to pay attention to aspects such as data preparation, model selection, and hyperparameter tuning. In this work, we focused on the influence of different featurization methods and algorithms on predicting the enthalpy of formation (EOF) of energetic compounds. We manually extracted a dataset consisting of 649 EOF values of energetic materials from the literature and compared different combinations of featurization methods and algorithms. The experimental results confirmed that ML can effectively map the relationship between molecular structure and EOF. Custom descriptor sets were found to perform best in featurization with a mean absolute error of 90.10 kJ mol(-1), after training by kernel ridge regression algorithm.