Predicting Material Strength Model Parameters Using a Machine Learning Based Approach

Machine learning (ML) is an emerging technology increasingly used to calibrate material models. While reported results often demonstrate good fit to experimental set, there remain concerns in the larger scientific community regarding overfitting of ML models to data and a lack of generalization to data outside that used for training. In this work, we investigate a problem using convolutional neural networks (CNNs) to calibrate the Preston-Tonks-Wallace (PTW) material strength model to hydrodynamic simulations of Richtmyer-Meshkov Instabilities (RMIs) for solid OFHC copper interfaces. Hundreds of parameterizations of the PTW material strength model are utilized within hydrodynamic simulations to generate synthetic PDV (Photon Doppler Velocimetry) data for RMI jet velocities, which are then used within the CNNs to learn the inverse mapping to the strength model parameters. We then investigate the sensitivities of these PDV quantities of interest to model parameters to better understand the generalization of the trained CNNs to data outside that used in the training process. This work demonstrates that the learning capability of neural networks for model parameterization correctly does not predict parameters that data is not sensitive to within the physical regime where the training data are sampled.