Fig. 7: Transfer learning applied to an independent experiment.

Shown are predictions from the cross-trained convolutional encoder-decoder (CED) model (experiment p4581) with normal loads that progressively increase from 3 to 8 MPa (see Fig. 6, (a-f), respectively). Each load level is predicted independently using the cross-trained model from simulation (the encoder and decoder) and data from experiment (p4677) conducted at 2.5 MPa. The predictions as manifest by the mean absolute percentage error (MAPE) progressively decrease with increasing load level. Nevertheless, the results show that the transfer learning approach with cross-training of the latent space, which accounts for only 20% of the total CED model parameters, is a powerful approach to predicting the frictional state of the experimental fault.