[1oral101-10] T1［Topic Session］Deformation and reaction of rocks and minerals

[T1-O-1] Deep Learning-based predicting of future frictional behavior via past physical parameters on experiment fault surface: A Transformer Architecture approach

*Tae-hoon UHMB¹, Takehiro Hirose², Yohei Hamada² (1. Disaster Prevention Research Center, Aichi Institute of Technology, 2. Kochi Institute for Core Sample Research, Japan Agency for Marin-Earth Science and Technology)

Keywords:future fault friction, PHV shear experiment, machine learning, Transformers Architecture

The fault friction is integral to earthquake physics, with friction playing a pivotal role in fault mechanics, such as earthquake initiation, energy dissipation during slip, and fault strength evolution. Therefore, the capability to predict future frictional behavior is a cornerstone in our quest to improve earthquake prediction and understanding. Yet, predicting future frictional behavior, a task crucial for substantial progress in earthquake prediction and understanding, remains unexplored due to the complexity of identifying relationships between past physical parameters and future friction. In this study, we constructed models, established by the Transformer Architecture, to predict future frictional behavior on experimental fault surfaces using past physical parameters, including normal stress, temperature, power density, axial shortening, and friction, recorded over specific timeframes (50 Hz). Transformers Architecture is a type of machine learning model, known for their 'attention' mechanism; they're widely used in language models like “ChatGPT” to generate natural languages. Our models are trained on sequential data obtained from the "Pressurized High-Velocity" (PHV) shear experiments conducted on Carrara Marble specimens. These experiments were performed at a constant slip velocity (0.01 m/s) under varying normal stress conditions ranging from 0.5 to 2.5 MPa. We initially evaluate the performance of each model in predicting future friction under varying slip conditions. Following this assessment, we identify the physical parameter that best elucidates future friction, based on the performance of the respective models. Further, we explore the temporal relationships between specific past physical parameter and future friction by analyzing the attention weights of the best-performing model. Our findings suggest that future friction behaviors under changing slip conditions can be predicted by certain physical parameters, notably past friction. Additionally, we observed that future friction at each given time is influenced by past physical parameters from each distinct time. We anticipate that these insights will establish a foundation for the development of future friction prediction models and contribute to a deeper understanding of fault friction behaviors.