Prediction of material properties of energetic materials using machine learning methods

The application of modern machine learning methods provides new possibilities for gaining information from existing data in many areas. In the case of energetic materials, such methods can help to gain a deeper understanding of the interaction of various influencing factors through exploratory data analysis, both for existing formulations and for the development of new ones. The prediction of specific material properties is also conceivable. The prerequisite for the application of machine learning methods is a set of high qualitative and quantitative data. For this purpose, a new extended data set was created based on an existing data set from the Fraunhofer ICT (Institute for Chemical Technology) Thermodynamics Database in combination with data retrieved from the public PubChem database. Several models were developed with this data set to predict the enthalpy of formation based on information about the composition of substances. Different methods were used for model development, ranging from simpler methods such as multivariate linear regression and partial least squares to more complex methods such as ridge regression and random forest, and even to deep neural networks. The following work describes an iteration of the development cycle of these models, starting with data collection and preprocessing, continuing with model building and hyperparameter tuning, and finally, evaluating and assessing the resulting models.