Machine Learning of Non-Reactive Shock Compression Response for Metallized Material Mixtures

The accurate prediction of the dynamic shock compression response of a material is crucial for a diverse range of engineering and scientific applications. Traditional methods, such as thermochemical and atomistic models, are powerful tools for predicting shock compression response. However, these methods are often expensive, time-consuming, and often rely on high-fidelity experimental data obtained after synthesizing the material and conducting expensive shock experiments. Consequently, such models have limited applicability to novel materials and speculative molecules or materials that may not exist. To address these issues, we need models that rapidly and accurately map from chemical species in composite mixtures to equations of state (EOS), which govern the shock compression response over a range of thermodynamic conditions.

In this study, we developed an ML framework to rapidly estimate the isentropic EOS of metallized mixtures. We used various ML models, such as multi- and single-task Gaussian Process Regression and Artificial Neural Networks, training it on a carefully curated dataset of 187 materials generated using EXPLO5. The trained ML models demonstrated a remarkable ability to generalize beyond the training data, with a high coefficient of determination, greater than 0.99 upon comparison with reference isentropic EOS curves. The study also examined the potential causes of variations in ML performance for some material classes.