9:30 AM - 9:45 AM
[SSS07-03] Data assimilation of seismic wavefields and estimation of source mechanism based on adjoint equations
Keywords:Seismology, Data assimilation, Adjoint method
1. Introduction
Recent advancements in dense seismic observations have led to the development of a new method for estimating earthquake sources. This method involves backpropagating observed waveforms towards the earthquake source to estimate the initial wavefield at the onset of the earthquake (McMechan, 1982; Larmat et al., 2006). It addresses issues encountered in conventional seismic source estimation, such as errors due to P/S wave picking accuracy, handling unclear microtremors, and separating simultaneous earthquakes. In this study, we aim for highly accurate source estimation, including absolute seismic amplitude, by introducing the adjoint method, a data assimilation technique, to seismic wave backpropagation calculations. Previous studies focused on predicting future shaking and tsunami propagation using seismic wave propagation simulations and observations with optimal interpolation method (Hoshiba & Aoki, 2015; Wang et al., 2017; Furumura et al., 2018). In contrast, our study utilizes the adjoint method to iteratively modify the wavefield to approach the real wavefield, resulting in higher accuracy compared to non-iterative methods such as optimal interpolation method. The effectiveness of the adjoint method has been demonstrated in numerical experiments for two-dimensional fields (Morita, Maeda, & Takano, 2023; Morita, Furumura, & Maeda, 2023). Building upon these results, we attempt to estimate the epicenter location and mechanism using real waveform data from dense seismic observations.
2. Methods and Data
We utilized waveform data from MeSO-net seismic stations, spanning the Kanto Plain from Doshimura, Yamanashi Prefecture, to Itako City, Ibaraki Prefecture, converted to velocity waveforms after applying a bandpass filter (0.05 to 0.5 Hz) (Fig. 1). Seismic wave backpropagation and forward propagation calculations were conducted in a two-dimensional domain (200 km horizontal x 100 km vertical) along MeSO-net's sidetracks, employing the JIVSM velocity structure model (Koketsu et al., 2012) (Fig. 2). Computational stability was ensured by setting the minimum S-wave velocity to 2.0 km/s for omitting shallow sedimentary layers. Seismic wave calculations utilized the staggered-grid finite difference method with appropriate spatial and temporal accuracies. Observed waveforms from the earthquake onset, including P-wave initial motion and S-wave, were backpropagated and forward propagated to obtain simulated waveforms at each observation point. Adjoint calculations iteratively corrected the wavefield until convergence, determining the source and earthquake onset time based on seismic wave convergence evaluated using elastic energy.
3. Results
We estimated the source of the December 29, 2021 earthquake beneath the 23 wards of Tokyo (depth: 31 km, Mj 3.5) using 31 MeSO-net stations spaced approximately 5 km apart. The calculations were performed using OpenMP parallel computing on a 1-CPU Wisteria supercomputer at the University of Tokyo's Information Technology Center, and took 2 hours for 2000 iterations. Results (Fig. 3) showed energy peaks near the estimated solutions of JMA and F-net, with good waveform agreement (VR=85%). The calculated mechanism closely resembled F-net's solution (Fig. 4). However, including the S-wave phase led to reduced VR (50%), likely due to the influence of the Kanto Plain's 3D sedimentary layer structure. Thinning observation points to 20 km intervals concentrated energy on sedimentary layers, not the source, but excessively high VR (98%) indicated convergence to a false epicenter.
4. Summary and Future prospects
Our study successfully estimated earthquake sources with high compatibility with existing solutions. However, 3D calculations are essential for regions like the Kanto Plain with large 3D heterogeneous structures, despite increased computational costs. Efforts to enhance computational efficiency, such as reducing iterations and improving imaging methods, are warranted.
5. Acknowledgments
We acknowledge computational support from Wisteria at the University of Tokyo's Center for Information Infrastructure and observational data from MeSO-net, National Research Institute for Earth Science and Disaster Prevention (NIED).
Recent advancements in dense seismic observations have led to the development of a new method for estimating earthquake sources. This method involves backpropagating observed waveforms towards the earthquake source to estimate the initial wavefield at the onset of the earthquake (McMechan, 1982; Larmat et al., 2006). It addresses issues encountered in conventional seismic source estimation, such as errors due to P/S wave picking accuracy, handling unclear microtremors, and separating simultaneous earthquakes. In this study, we aim for highly accurate source estimation, including absolute seismic amplitude, by introducing the adjoint method, a data assimilation technique, to seismic wave backpropagation calculations. Previous studies focused on predicting future shaking and tsunami propagation using seismic wave propagation simulations and observations with optimal interpolation method (Hoshiba & Aoki, 2015; Wang et al., 2017; Furumura et al., 2018). In contrast, our study utilizes the adjoint method to iteratively modify the wavefield to approach the real wavefield, resulting in higher accuracy compared to non-iterative methods such as optimal interpolation method. The effectiveness of the adjoint method has been demonstrated in numerical experiments for two-dimensional fields (Morita, Maeda, & Takano, 2023; Morita, Furumura, & Maeda, 2023). Building upon these results, we attempt to estimate the epicenter location and mechanism using real waveform data from dense seismic observations.
2. Methods and Data
We utilized waveform data from MeSO-net seismic stations, spanning the Kanto Plain from Doshimura, Yamanashi Prefecture, to Itako City, Ibaraki Prefecture, converted to velocity waveforms after applying a bandpass filter (0.05 to 0.5 Hz) (Fig. 1). Seismic wave backpropagation and forward propagation calculations were conducted in a two-dimensional domain (200 km horizontal x 100 km vertical) along MeSO-net's sidetracks, employing the JIVSM velocity structure model (Koketsu et al., 2012) (Fig. 2). Computational stability was ensured by setting the minimum S-wave velocity to 2.0 km/s for omitting shallow sedimentary layers. Seismic wave calculations utilized the staggered-grid finite difference method with appropriate spatial and temporal accuracies. Observed waveforms from the earthquake onset, including P-wave initial motion and S-wave, were backpropagated and forward propagated to obtain simulated waveforms at each observation point. Adjoint calculations iteratively corrected the wavefield until convergence, determining the source and earthquake onset time based on seismic wave convergence evaluated using elastic energy.
3. Results
We estimated the source of the December 29, 2021 earthquake beneath the 23 wards of Tokyo (depth: 31 km, Mj 3.5) using 31 MeSO-net stations spaced approximately 5 km apart. The calculations were performed using OpenMP parallel computing on a 1-CPU Wisteria supercomputer at the University of Tokyo's Information Technology Center, and took 2 hours for 2000 iterations. Results (Fig. 3) showed energy peaks near the estimated solutions of JMA and F-net, with good waveform agreement (VR=85%). The calculated mechanism closely resembled F-net's solution (Fig. 4). However, including the S-wave phase led to reduced VR (50%), likely due to the influence of the Kanto Plain's 3D sedimentary layer structure. Thinning observation points to 20 km intervals concentrated energy on sedimentary layers, not the source, but excessively high VR (98%) indicated convergence to a false epicenter.
4. Summary and Future prospects
Our study successfully estimated earthquake sources with high compatibility with existing solutions. However, 3D calculations are essential for regions like the Kanto Plain with large 3D heterogeneous structures, despite increased computational costs. Efforts to enhance computational efficiency, such as reducing iterations and improving imaging methods, are warranted.
5. Acknowledgments
We acknowledge computational support from Wisteria at the University of Tokyo's Center for Information Infrastructure and observational data from MeSO-net, National Research Institute for Earth Science and Disaster Prevention (NIED).