[HCG28-P05] Machine learning for activity evaluation of crush zones using chemical composition: introduction of examples
Keywords:machine learning, crush zone, chemical composition
We need a method to assess the activity of crush zones alternative to the application of overlying sediments (Hataya et al., 1992), because the zones encountered in underground tunnels and boring have unknown extensions to the ground surface. Until now, a cease of crush zone activity has been proven by the existence of (a) disagreement of current stress field and slip direction, (b) not-destroyed dike, vein, minerals on the basis of cross-cutting relation, and (c) high-temperature condition of deformation microstructures or mineralization generated when these incidents occurred accompanied with water-rock interaction (NRA, 2015; Ishiwatari, 2016; Niwa et al., 2016; Shimada et al., 2016). These methods have been advanced as being applicable in the field, but ambiguity or inconsistency of evaluation, even among experts, are exist. So, there is room for further evaluation method development.
The method to be developed is one in which the result is objective and independent of the person, which helps professional judgment. In addition, implementation, dissemination and verification must be executable by a general geological engineer. Ideally, it should be as easy as possible to be understood by non-experts. In light of these goals, the whole-rock chemical composition of the fault gouge along a principal slip zone of the crush zone is attractive. In underground tunnels, the time for geological observation may be limited due to construction and safety measures, whereas sampling for whole-rock chemical composition can be done in a relatively short time. Considering the case of widespread use in the future, the hurdle of sampling is low even when a sample is provided by a volunteer. Furthermore, there is a possibility that non-destructive in-situ measurement can be performed by portable XRF. Chemical analysis data is quantitative data that can be disclosed in literatures, and the database can be continuously expanded and repeatedly verified. Then, is there any difference in the chemical composition of fault rocks between active and inactive faults? Is there a solution worth asking this question?
Kuwatani et al. (2014) show that the chemical composition of tsunami deposits and non-tsunami deposits from the 2011 off the Pacific coast of Tohoku Earthquake can be separated with high probability by machine learning (multivariate analysis). We thought this idea could be a solution. Therefore, following this concept, we collected literature values of the chemical composition of fault gouges for faults with known activities, and began searching for a primary equation that distinguishes active from non-active faults by multivariate analysis in 2018. The explanatory variable is a chemical composition which is a quantitative variable, and the objective variable to be predicted is a fault property (active fault or non-active fault) which is a qualitative variable. See Tateishi et al. (2020) for details of the results. The results of studies on granitic rocks show that there are multiple discriminants that separate active and non-active faults with a discrimination rate of 100%. In the process of searching for a discriminant with high generalization performance that is not specialized for teacher data, elements that greatly contribute to discrimination between the two groups are being found. In other words, it is an element that characterizes active faults, and clues are being obtained for solving the nature of chemical composition change associated with continuous activity or activity cessation. In the presentation, we would like to introduce the current status of initiatives, including past case studies.
This study was funded by METI, Japan as part of its R&D program supporting development of technology for geological disposal of HLW.
References are shown in Japanese abstract.
The method to be developed is one in which the result is objective and independent of the person, which helps professional judgment. In addition, implementation, dissemination and verification must be executable by a general geological engineer. Ideally, it should be as easy as possible to be understood by non-experts. In light of these goals, the whole-rock chemical composition of the fault gouge along a principal slip zone of the crush zone is attractive. In underground tunnels, the time for geological observation may be limited due to construction and safety measures, whereas sampling for whole-rock chemical composition can be done in a relatively short time. Considering the case of widespread use in the future, the hurdle of sampling is low even when a sample is provided by a volunteer. Furthermore, there is a possibility that non-destructive in-situ measurement can be performed by portable XRF. Chemical analysis data is quantitative data that can be disclosed in literatures, and the database can be continuously expanded and repeatedly verified. Then, is there any difference in the chemical composition of fault rocks between active and inactive faults? Is there a solution worth asking this question?
Kuwatani et al. (2014) show that the chemical composition of tsunami deposits and non-tsunami deposits from the 2011 off the Pacific coast of Tohoku Earthquake can be separated with high probability by machine learning (multivariate analysis). We thought this idea could be a solution. Therefore, following this concept, we collected literature values of the chemical composition of fault gouges for faults with known activities, and began searching for a primary equation that distinguishes active from non-active faults by multivariate analysis in 2018. The explanatory variable is a chemical composition which is a quantitative variable, and the objective variable to be predicted is a fault property (active fault or non-active fault) which is a qualitative variable. See Tateishi et al. (2020) for details of the results. The results of studies on granitic rocks show that there are multiple discriminants that separate active and non-active faults with a discrimination rate of 100%. In the process of searching for a discriminant with high generalization performance that is not specialized for teacher data, elements that greatly contribute to discrimination between the two groups are being found. In other words, it is an element that characterizes active faults, and clues are being obtained for solving the nature of chemical composition change associated with continuous activity or activity cessation. In the presentation, we would like to introduce the current status of initiatives, including past case studies.
This study was funded by METI, Japan as part of its R&D program supporting development of technology for geological disposal of HLW.
References are shown in Japanese abstract.