Résumé
We present a new model for the fragmentation behavior of granular materials based on Artificial NeuralMaps (ANMs), a deep learning architecture. To generate learning datasets, we consider the fragmentationof a unit square including a controlled level of disorder subjected to diametral traction and simulatedby the Peridynamic method. More than 900 samples were simulated and the output data recorded upto failure. We used a specific 3-layer autoencoder for each pixel described by its fracture state (failed orunfailed). The loss function is based on the correlation between the dataset and the prediction. When anotch is introduced, the cracks are initiated from the notch and the crack is predicted with an accuracyof 100% independently of disorder. The speedup is at least 10000-fold between the peridynamic andANM-based simulations. Our results show that ANMs can accurately capture the complex crack paths inthe presence of disorder; see Figure. ANMs can be easily implemented in a DEM software to simulate anassembly of breakable particles with built-in texture.