A Machine Learning Framework to Predict Unreacted Shock Compression Response of Solids Under Scarce Data

The design of new materials based on shock-compressed properties is becoming increasingly important. However, in scenarios where a rapid evaluation of many candidate materials is needed, current approaches for measuring or simulating unreacted shock properties are unsuitable due to the attendant high cost. To this end, we examine machine learning (ML) techniques to estimate the shock-Hugoniot curves of a diverse set of materials. We perform systematic model evaluation and feature engineering to examine approaches for overcoming challenges associated with scarce data. A novel featurization strategy combining physical and synthetic features is developed and implemented showing substantial improvements in ML prediction accuracy of complete shock-Hugoniot curves. The formal study performs exhaustive evaluation to analyze outlier predictions, to understand the effects of variance in source data on prediction performance, and additional data fusion and knowledge sharing strategies to enable ML for scarce data. The emergence of new ML algorithms will have wide-ranging impacts on ML-assisted materials design under shock conditions.