Machine learning boosted first principles model predictions of <scp>DSC</scp> decomposition energies for early‐stage material screening

Abstract Identification of high‐energy compounds used in the pharmaceutical industry has been made easy via differential scanning calorimetry, but when designing new synthetic routes, working with new or limited amounts of material, or the inability to isolate an intermediate, differential scanning calorimetry and other thermal hazard data may not be readily available. Here we report a machine learning model that uses a first principles model as a baseline for predicting the decomposition energies of materials without having to know the decomposition products beforehand. The first principles model depends on bond dissociation energies to simulate bond breaking and then subsequent reconstruction of various likely decomposition products, summing the energy consumed and released. A light gradient boosting machine model was trained using the first principles model along with several other features and showed a significant improvement in accuracy compared with another trained model omitting the first principles model, especially for molecules without any obvious high energy functional groups. Another two models were trained to predict shock sensitivity and explosivity potential with high accuracy. These first principles boosted machine learning models allow for informed decisions to be made regarding the thermal hazards of materials before synthesizing, isolating, or purchasing them.