Convolutional LSTM surrogate for mesoscale hydrocode simulations of granular wave propagation

Abstract

Granular materials subjected to impact loading exhibit heterogeneous spatiotemporal dynamics governed by stress-wave propagation, pore collapse, and grain-scale rearrangement. Mesoscale hydrocodes can resolve these mechanisms, but the computational cost of repeated simulations limits parametric studies, uncertainty quantification, and rapid screening. We develop a convolutional Long Short-Term Memory (ConvLSTM) neural network as an image-sequence surrogate for rendered pressure-field outputs from two-dimensional FLAG mesoscale hydrocode simulations. The model is trained using overlapping sliding windows from available simulation trajectories and is evaluated on simulations withheld at the trajectory level. We first consider a controlled “billiard break” problem in which a deformable cue ball impacts a cluster of nine deformable circular balls. The trained surrogate reproduces the dominant wavefront evolution and particle motion for withheld combinations of cue-ball position and velocity. We then train a separate ConvLSTM model using the same architecture and training procedure for weak-shock compaction of a granular ensemble, evaluating piston speeds withheld from the granular training set. The surrogate captures the location and broad morphology of the compaction front, while smoothing fine pore-scale features in compacted regions. The present model should be interpreted as a surrogate for rendered RGB pressure-field composites rather than a conservative surrogate for the full hydrocode state. These results establish a baseline data-driven image-sequence framework for accelerating exploration of mesoscale granular wave dynamics and identify the key extensions required for physical-field prediction, frictional systems, and hot-spot-sensitive applications.