10:15 AM - 10:30 AM
[SCG51-06] Objective Estimation of Mechanical Plate Boundary by Neural Network
Keywords:Neural network, Plate boundary, Focal mechanism
It is critically important in the dynamics of plate subduction, which is a fundamental topic in Earth science (Becker and Faccenna, 2009), to determine plate interfaces. Since either erosion or underplating occurs at the plate interface, there are two types of boundaries: material and mechanical (Shreve and Cloos, 1986). Most previous approaches determined material plate boundaries (Zhao et al., 1997). Meanwhile, several studies estimated physical plate interfaces (Nakajima and Hasegawa, 2006) based on earthquake selection and assumptions, which is not necessarily objective. Hence, we seek a new method to estimate mechanical plate boundaries objectively.
This study proposes an approach to estimate mechanical plate boundaries using a neural network to focal mechanism database, inspired by recent advances of machine learning in Earth science (Asim et al., 2017; Mousavi et al., 2019). Specifically, our neural network predicts focal mechanisms from hypocentral information. We used the first-motion focal mechanisms determined by JMA from 1997 October through 2018 December. As a pre-processing, we convert the focal mechanism information to the squared sine of plunges for P-, N-, and T-axes, which corresponds to the coordinate in the triangle diagram of Frohlich (1992). Then, we divide the data randomly into 80% for learning and 20% for validation. We define loss function as mean square error and use Adam for optimization. The network architecture is also optimized using Optuna, resulting in four hidden layers, each with 428 nodes.
The advantage of neural network modeling is estimating the focal mechanism, or stress state, at an arbitrary point by adaptively interpolating the data. First, therefore, we estimate the focal mechanisms near the surface using the neural network and confirm their consistency with geological fault survey studies (Headquarters of Earthquake Research Promotion). The final remaining loss in the learning process might be due to the worse performance in the area where different fault types co-exist.
Next, we estimate mechanism type across the plate subduction of the Pacific plate in Tohoku and Hokkaido areas. We determine the mechanical plate boundary depth by the most significant reverse-fault component between the previously given depth and 30 km deeper. As a result, the mechanical plate boundary at the depths around 50 km is 5–10 km deeper than the previous study. Assuming that the plate interface of the previous study is a material boundary, our results imply the underplating on the Pacific plate. We need further tests to evaluate the contribution of hypocentral precision.
At these depths, the physical plate interface has roughness at km-scale. Considering that the focal mechanism at the region typically has a dip angle (20–40º) larger than the typical slope of plate interface (~10º), we interpret that echelon structure, which we often observe at strike-slip faults (Sibson, 1986), also exist in vertical profiles. Such a rough plate interface is consistent with previous studies (Zhan et al., 2012; Boneh, 2019). The mechanical plate interface at the depths of 100–200 km has larger-scale steps of tens of kilometers, although it might be an artifact considering that the number of deep events is small and that one of the input layers is for depth.
In summary, we designed a neural network for focal mechanism data and estimated the mechanical plate boundary objectively. We expect such an approach could also be applied for stress inversion.
This study proposes an approach to estimate mechanical plate boundaries using a neural network to focal mechanism database, inspired by recent advances of machine learning in Earth science (Asim et al., 2017; Mousavi et al., 2019). Specifically, our neural network predicts focal mechanisms from hypocentral information. We used the first-motion focal mechanisms determined by JMA from 1997 October through 2018 December. As a pre-processing, we convert the focal mechanism information to the squared sine of plunges for P-, N-, and T-axes, which corresponds to the coordinate in the triangle diagram of Frohlich (1992). Then, we divide the data randomly into 80% for learning and 20% for validation. We define loss function as mean square error and use Adam for optimization. The network architecture is also optimized using Optuna, resulting in four hidden layers, each with 428 nodes.
The advantage of neural network modeling is estimating the focal mechanism, or stress state, at an arbitrary point by adaptively interpolating the data. First, therefore, we estimate the focal mechanisms near the surface using the neural network and confirm their consistency with geological fault survey studies (Headquarters of Earthquake Research Promotion). The final remaining loss in the learning process might be due to the worse performance in the area where different fault types co-exist.
Next, we estimate mechanism type across the plate subduction of the Pacific plate in Tohoku and Hokkaido areas. We determine the mechanical plate boundary depth by the most significant reverse-fault component between the previously given depth and 30 km deeper. As a result, the mechanical plate boundary at the depths around 50 km is 5–10 km deeper than the previous study. Assuming that the plate interface of the previous study is a material boundary, our results imply the underplating on the Pacific plate. We need further tests to evaluate the contribution of hypocentral precision.
At these depths, the physical plate interface has roughness at km-scale. Considering that the focal mechanism at the region typically has a dip angle (20–40º) larger than the typical slope of plate interface (~10º), we interpret that echelon structure, which we often observe at strike-slip faults (Sibson, 1986), also exist in vertical profiles. Such a rough plate interface is consistent with previous studies (Zhan et al., 2012; Boneh, 2019). The mechanical plate interface at the depths of 100–200 km has larger-scale steps of tens of kilometers, although it might be an artifact considering that the number of deep events is small and that one of the input layers is for depth.
In summary, we designed a neural network for focal mechanism data and estimated the mechanical plate boundary objectively. We expect such an approach could also be applied for stress inversion.