Convolutional neural network-based moment tensor inversion using domain adaptation for microseismicity monitoring

Abstract

Microseismic monitoring is widely used to analyze the locations and growth directions of fractures formed at sites of hydraulic fracturing treatment and CO2 geologic sequestration. Because moment tensors can provide focal mechanisms, moment tensor inversion has received considerable attention in microseismic monitoring; the real-time processing of moment tensor inversion is important for rapid decision-making. Pre-trained machine learning (ML) models can make nearly instantaneous predictions in the application stage and thus present an attractive alternative to real-time processing. However, prior information regarding the velocity model at the target site is a prerequisite for generating the dataset used to train the ML model that is applied in moment tensor inversion. In addition, it is difficult to create the training dataset because it requires three-dimensional numerical modelling when the velocity model is complex; numerous simulations must be executed for sources with various locations and moment tensors. To overcome these limitations, we applied the domain adaptation technique to the convolutional neural network (CNN)-based moment tensor inversion method, which uses peak amplitudes and arrival times of P- and S-waves as input features. The CNN model was pre-trained with the dataset generated from a homogeneous velocity model. Then, in the domain adaptation stage, the pre-trained model was fine-tuned along with the target dataset. To validate the performance of the domain adaptation, moment tensors from both horizontal and tilted three-layer models were predicted. In each case, the domain-adapted model performance was similar to the performances of the CNN-based models that had been trained using the dataset generated with the exact target velocity models.