Verification of a Machine Learning Model through Quantum Chemical Simulation

The development of new formulations of energetic materials is a costly and time-consuming process. In the search for ideal material combinations to achieve the desired target parameters of a new propellant or explosive, the cycle of design, synthesis, characterization, and experimental verification is run through several times. In the design phase, simulation models and tools of varying complexity from the macroscopic to the quantum mechanical level can be used to obtain predictions about specific properties of single components. Another somewhat newer method that is becoming increasingly popular is the prediction of material properties using adapted machine learning models. As part of a previous work a model for predicting the enthalpy of formation was created and presented at last year's 52nd International Annual Conference of the Fraunhofer ICT. This model has since been further developed and optimized. The next development step is now the verification of the existing model. As part of this verification, the enthalpies of formation of some molecules were determined by quantum chemical simulation and compared with the predictions of the machine learning model. The adaptation and the results of both methods are described in the following work.