09:15 〜 09:30
[SSS07-14] What do the time-variable friction parameters in laboratory experiments tell us about sliding dynamics?
Rock friction is an essential part of the sliding dynamics on earthquake faults. The Rate- and State-dependent Friction (RSF) law (Dieterich, 1978, 1979) is widely used as a constitutive law for rock friction. However, the physical basis of the RSF law and its friction parameters are not well understood yet. In previous studies, the friction parameters were related to the gouge layer thickness (e.g., Beeler et al., 1996) and it was reported that this relationship also holds for the stick-slip events observed in the large-scale rock friction experiments (Urata et al., 2017).
The goal of this research is to estimate the time evolution of friction parameters and discuss the relationship between these parameters and fault properties such as the gouge layer thickness. To get a more comprehensive understanding of their temporal changes, we estimated these parameters for hundreds of stick-slip events occurring on laboratory faults, whereas these parameters were estimated for just a few stick-slip events in Urata et al. (2017). Since the amount of data is large, we adopted a Random Forest (RF) machine learning approach (Breiman, 2001), which has been widely used in the recent years to analyze geophysical data (e.g., Rouet-Leduc et al., 2017).
The RF model was trained on simulated friction data and then applied to the stick-slip events of experiment data (Fukuyama et al., 2014) to estimate the friction parameters. The gouge particles on the faults were removed before the experiments (Fukuyama et al., 2014), which made it possible to evaluate the influence of the evolution of the gouge layer on the variation of friction parameters. When generating the synthetic friction data for stick-slip events, a one-degree-of-freedom mass-spring system with RSF law was assumed. The estimation procedure was as follows: 1) We computed the friction behavior of stick slip events for various values of a, b-a and Lc parameters. 2) We selected some ‘representative features’ in the simulated data for each combination of parameters. 3) The RF model was trained using the simulated data with the parameter sets with respect to the corresponding representative features. 4) The representative features for each observed stick-slip event were also measured from the experiment data. 5) These representative features were used as input of the trained RF model to estimate the friction parameters for the experiment data.
Using the RF approach, we captured the temporal variation of the friction parameters a, b-a and Lc. The time required to estimate the parameters for a single stick-slip event was one twentieth compared to that of Urata et al. (2017). No significant problem such as overfitting, which is sometimes reported when applying machine learning, was observed throughout the estimation. The temporal variation of friction parameters may be related with the evolution of the gouge layer: during a first transient phase, the parameter a becomes smaller, while b-a and Lc become larger, as the gouge layer becomes thicker. The variation of the parameters becomes less pronounced during the following steady-state phase. Further research on the evolution of friction parameters, gouge layer and other properties of faults, such as temperature and slip rate, may help us understand the underlying physical processes of the RSF law.
The goal of this research is to estimate the time evolution of friction parameters and discuss the relationship between these parameters and fault properties such as the gouge layer thickness. To get a more comprehensive understanding of their temporal changes, we estimated these parameters for hundreds of stick-slip events occurring on laboratory faults, whereas these parameters were estimated for just a few stick-slip events in Urata et al. (2017). Since the amount of data is large, we adopted a Random Forest (RF) machine learning approach (Breiman, 2001), which has been widely used in the recent years to analyze geophysical data (e.g., Rouet-Leduc et al., 2017).
The RF model was trained on simulated friction data and then applied to the stick-slip events of experiment data (Fukuyama et al., 2014) to estimate the friction parameters. The gouge particles on the faults were removed before the experiments (Fukuyama et al., 2014), which made it possible to evaluate the influence of the evolution of the gouge layer on the variation of friction parameters. When generating the synthetic friction data for stick-slip events, a one-degree-of-freedom mass-spring system with RSF law was assumed. The estimation procedure was as follows: 1) We computed the friction behavior of stick slip events for various values of a, b-a and Lc parameters. 2) We selected some ‘representative features’ in the simulated data for each combination of parameters. 3) The RF model was trained using the simulated data with the parameter sets with respect to the corresponding representative features. 4) The representative features for each observed stick-slip event were also measured from the experiment data. 5) These representative features were used as input of the trained RF model to estimate the friction parameters for the experiment data.
Using the RF approach, we captured the temporal variation of the friction parameters a, b-a and Lc. The time required to estimate the parameters for a single stick-slip event was one twentieth compared to that of Urata et al. (2017). No significant problem such as overfitting, which is sometimes reported when applying machine learning, was observed throughout the estimation. The temporal variation of friction parameters may be related with the evolution of the gouge layer: during a first transient phase, the parameter a becomes smaller, while b-a and Lc become larger, as the gouge layer becomes thicker. The variation of the parameters becomes less pronounced during the following steady-state phase. Further research on the evolution of friction parameters, gouge layer and other properties of faults, such as temperature and slip rate, may help us understand the underlying physical processes of the RSF law.