[S-CG45] Science of slow-to-fast earthquakes

[SCG45-P04] An experiment-based fault friction modeling: An approach applying Recurrent neural network (RNN)

*オム テフン¹、濱田 洋平¹、廣瀬 丈洋¹ (1.国立研究開発法人海洋研究開発機構　高知コア研究所)

キーワード：地震、摩擦法則、摩擦予測、機械学習

Friction is a important factor in understanding earthquake physics. Earthquakes are considered as a frictional instability of a pre-existing fault or plate interface, and the frictional behavior of fault rocks is controlled by the friction law of rock. The friction laws have been deeply studied for centuries on the basis of micromechanics model of contact and Rate and State Theory, trying to explain the relationships between friction and physical parameters (e.g., velocity, temperature, surface chemistry, material properties and so on) as mathematical model (Marone, 1998; Aharonov and Scholz, 2018). However, there are still challenges in applying the classical model to define friction due to complex interdependences of the physical parameters and evolution of fault surface with slip time. Here we developed a new model defining friction of experimental fault for various physical parameters (velocity, normal stress, temperature, temperature rate, displacement normal to slip surface, rate of displacement normal to slip surface) and slip time by applying Recurrent neural networks (RNNs) to experiment data reported previously from Hirose et al., (2012). RNNs are a class of machine learning techniques that employ a memory mechanism to process sequential data. The experiment has been conducted on diorite specimens at a constant slip velocity (0.004 m/s) in various normal stress (0.7-3.8 MPa), using a rotary-shear, high speed friction apparatus (HVR 1087). First, we identified an optimal model through an assessment of its performance in predicting friction with changing slip conditions (normal stress, temperature, temperature rate, displacement normal to slip surface, rate of displacement normal to slip surface). Then, we investigated relationships between friction and the physical parameters, including the inter-parameter correlations by analyzing gradients for the parameters of the optimal model. Moreover, we interpreted time-dependences of the correlations for friction. Our results indicate that friction behavior with changing slip conditions can be predicted by the physical parameters and that the importance of these parameters on friction are shifting with changing slip conditions and time. We expect our results to serve as a foundation for the development of more sophisticated friction model, as well as for predicting the future state of fault friction.