Supervised domain adaptation in prediction of peak shear strength of rock fractures

It is of great importance to determine peak shear strength (PSS) of rock fractures, and data-driven criteria have showed advances in fitting capability in recent years. However, the generalization ability of existing data-driven criteria is limited by dataset size and fracture roughness characterization, which is negative to predictive power and robustness of models. Here we proposed a novel data-driven criterion to predict PSS of rock fractures, with high generalization ability on real experimental data. We first created large-scale low-fidelity dataset by discrete-element modeling, and small-scale high-fidelity dataset by laboratory direct shear tests. The numeric features include normal stress, mechanical properties (including PSS of intact and flat-fracture rock specimens), secondary properties (including internal friction angle, cohesion strength and basic friction angle), and the matrixed feature is topography data. We then established domain adaptation (DA) models for cross-domain knowledge transfer between the low- and high-fidelity datasets, and roughness features were automatically extracted by convolution kernels. The best DA-based model is weighting adversarial neural network, outranking other models by error indicator, and the average relative error on experimental data of new rock types is within 10.0 %. Finally, the sensitivity of input features is investigated, which further proves the promising potential of the developed data-driven PSS criterion of rock fractures in engineering practice.