10:45 〜 11:00
[SCG55-01] 深層学習によるマルチチャンネル反射法地震探査データの速度モデル構築の高度化
キーワード：マルチチャンネル地震探査、深層学習、畳み込みニューラルネットワーク
At the Nankai Trough margin, the great megathrust earthquakes have generated strong motions and large tsunamis. The plate-boundary fault (decollement) is a source fault of the devastating Nankai megathrust earthquakes. There have been several seismic reflection studies conducted across the Nankai Trough margin to reveal the seismic structures of the Nankai Trough.
Although it is extremely important to construct a reliable velocity model for the seismic survey line and to calculate the physical properties of the subducting sediments, only a few studies on pore-fluid pressure estimation were conducted. One of the hurdles to accomplish this goal has been determining a P-wave velocity model from multi-channel seismic (MCS) reflection data, which takes a huge amount of time. Therefore, a more efficient approach is required to handle the large volume of MCS data along the Nankai Trough subduction zone. Based on these reasons, we are developing a method that can estimate the seismic velocity model, and eventually pore-fluid pressure, along the plate-boundary faults faster and more accurately by using deep learning.
In this research, we aim to incorporate convolutional neural networks (CNN) into MCS data processing and use the well-trained model to build P-wave velocity models faster. In the first step of our work, we used the Marmousi2 model (Fig. 1a) to generate the dataset for training our neural network. We first performed forward modeling using the Marmousi2 model to generate synthetic seismic dataset and then applied pre-stack depth migration (PSDM) on the synthetic data using a constant water velocity of 1500 (m/s). The common image gathers of the migrated data (with the sub-optimal water velocity) are used to calculate the vertical semblance panels for residual velocity analysis (Fig. 1b). The semblance panel (as the input) and vertical profile of the true interval velocity (as the response) for each CDP are used as pairs of training dataset. The whole dataset is randomly shuffled and split to training, validation, and test datasets. In the neural network training step, we randomly initialize the weights of the network and train the model with the shuffled dataset. After 200 training epochs, we examine the neural network with test dataset, which has not been included in the training set before. The network can predict the velocity profiles from Marmousi2 model within 100 m/s error (Fig. 1c).
Here, the neural network model was trained with synthetic data, however more diverse synthetic datasets, and eventually real data will be used to improve the model in the future. In the final step of our work, we aim to calculate the physical properties and structure of the decollement over a wide area of the Nankai Trough.
Although it is extremely important to construct a reliable velocity model for the seismic survey line and to calculate the physical properties of the subducting sediments, only a few studies on pore-fluid pressure estimation were conducted. One of the hurdles to accomplish this goal has been determining a P-wave velocity model from multi-channel seismic (MCS) reflection data, which takes a huge amount of time. Therefore, a more efficient approach is required to handle the large volume of MCS data along the Nankai Trough subduction zone. Based on these reasons, we are developing a method that can estimate the seismic velocity model, and eventually pore-fluid pressure, along the plate-boundary faults faster and more accurately by using deep learning.
In this research, we aim to incorporate convolutional neural networks (CNN) into MCS data processing and use the well-trained model to build P-wave velocity models faster. In the first step of our work, we used the Marmousi2 model (Fig. 1a) to generate the dataset for training our neural network. We first performed forward modeling using the Marmousi2 model to generate synthetic seismic dataset and then applied pre-stack depth migration (PSDM) on the synthetic data using a constant water velocity of 1500 (m/s). The common image gathers of the migrated data (with the sub-optimal water velocity) are used to calculate the vertical semblance panels for residual velocity analysis (Fig. 1b). The semblance panel (as the input) and vertical profile of the true interval velocity (as the response) for each CDP are used as pairs of training dataset. The whole dataset is randomly shuffled and split to training, validation, and test datasets. In the neural network training step, we randomly initialize the weights of the network and train the model with the shuffled dataset. After 200 training epochs, we examine the neural network with test dataset, which has not been included in the training set before. The network can predict the velocity profiles from Marmousi2 model within 100 m/s error (Fig. 1c).
Here, the neural network model was trained with synthetic data, however more diverse synthetic datasets, and eventually real data will be used to improve the model in the future. In the final step of our work, we aim to calculate the physical properties and structure of the decollement over a wide area of the Nankai Trough.