[S-CG40] Science of slow-to-fast earthquakes

[SCG40-15] Predicting of future frictional behaviors on experiment fault surfaces: A Transformer Architecture approach

★Invited Papers

*Tae-Hoon Uhmb¹, Yohei Hamada², Takehiro Hirose² (1.Disaster Prevention Research Center, Aichi Institute of Technology, 2.Japan Agency for Marine-Earth Science and Technology Kochi Institute for Core Sample Research)

Keywords:Fault Mechanics, Experimental Fault Friction Forecasting, Transformer Architecture, Natural Language Models

Friction is a key factor in determining the mechanics of faults, influencing the initiation of earthquakes, energy release during sliding events, and the development of fault strength. In a sense, forecasting frictional behavior is essential for gaining insights into the intricate dynamics of earthquake science. Although numerous studies have been conducted to comprehend fault friction, unraveling the link between future fault friction and historical physical parameters on the fault surface remains challenging, mainly due to difficulties in establishing these connections. In this research, we introduce novel models designed to forecast future friction using historical physical data from experimental fault surfaces, such as power density, friction, temperature, axial displacement, and normal stress. These models utilize the Transformer architecture, initially introduced for building large natural language model such as “ChatGPT” and “Bard”, and adapt it for analyzing sequential data in a rotary-shear experiment using Carrara marble samples. The experiment was conducted at a steady slip velocity of 0.01 m/s, with normal stress varying between 0.5 and 2.5 MPa over two cycles. We identify which parameters most significantly influence future friction by evaluating performances of the models, that apply each different past physical parameter, for predicting future friction. Subsequently, we elucidated the relationships between future friction and the most predictive past parameters under varying slip times and conditions, analyzing attention weights within the models. Our findings reveal that past friction is most closely linked to future friction in our architecture, and this relationship dynamically evolves with changing slip time and conditions. Our study underscores the value of deep learning in enhancing our understanding of fault physics and improve predictions of fault friction, contributing to the broader field of earthquake research.