Convolutional LSTM Surrogate for Mesoscale Hydrocode Simulations of Granular Wave Propagation
Granular materials subjected to impact loading exhibit highly heterogeneous spatiotemporal dynamics governed by wave propagation, pore collapse, and grain-scale rearrangements. Mesoscale hydrocodes resolve these processes but are computationally expensive, limiting their use in parametric studies and uncertainty quantification. In this work, we develop a convolutional Long Short-Term Memory (ConvLSTM) neural network as a spatiotemporal surrogate for mesoscale simulations of weak shock propagation in granular media. Using two-dimensional hydrocode simulations as training data, we first consider a simplified "billiard break" problem in which a cue ball impacts a cluster of nine circular balls, all deformable. Sequences of pressure-field images serve as input-output pairs for a sequence-to-sequence ConvLSTM, which is trained to predict future frames from a short history. We compare several architectures and show that a relatively compact encoder-decoder ConvLSTM accurately reproduces the propagation of the pressure wave and the resulting particle motion for an unseen combination of cue-ball position and impact velocity. As a proof-of-concept extension, we apply the same ConvLSTM framework to previously published mesoscale simulations of weak shock compaction in a granular ensemble. When evaluated at piston impact speeds that were completely withheld from training, the surrogate captures the position and shape of the compaction front and its dependence on impact speed, while smoothing fine pore-scale details in the highly compacted region as expected. These results demonstrate that ConvLSTM models can serve as satisfactory surrogates for spatiotemporal mesoscale simulations of granular wave propagation, enabling accelerated exploration of parameter space and laying the groundwork for physics-informed, mesoscale simulations of granular materials under shock loading.