Fig. 1: Workflow with numerical simulation (subfigure adapted from ref. ²²) and experimental configurations (subfigure adapted from ref. ⁶) for the transfer learning analysis.

See Methods section for full details about the simulation and experimental data. From the simulation and experiment, characteristics are obtained that include the shear displacement, shear stress, normal stress, bulk friction, kinetic energy (simulation), and acoustic emission (experiment). The simulation kinetic energy is used as a proxy for acoustic emission to train the model to predict the instantaneous bulk friction coefficient at all times throughout the slip cycle. Only the model latent space is then further trained using limited acoustic emission data from the laboratory experiment (number p4677). The simulation-trained encoder and decoder are left unchanged. The new model is used to predict the instantaneous friction for experimental data the model has not previously seen, from a different laboratory experiment (p4581). FDEM finite-discrete element method, CED convolutional encoder-decoder.