Sensitivity analysis for successful microseismic moment tensor inversion using machine learning

Abstract

As development of machine learning technology, many researchers are trying to graft machine learning into seismic processing. However, it is hard to create the training dataset for machine learning (ML) model used in moment tensor inversion because it needs 3D numerical modeling considering source moment tensor with complex velocity model. In order to efficiently make training dataset for ML model, it is necessary to analyze the factors affecting the performance of ML model. In this study, we constructed a 2D convolutional neural network (CNN) model that predicts moment tensor components of the microseismic event using peak amplitudes and traveltimes of P and S waves as input features. To examine the factors affecting the results of moment tensor inversion using ML model, we analyze the sensitivity of three factors, the error in the peak amplitude due to noise, the error in travel time due to picking error and the difference between the velocity model used in creating training data set and the true velocity model of the target area. The results show that the ML model for moment tensor inversion is most affected by the error in the peak amplitude. For the picking error, proper results are obtained up to the error by half wavelength. In addition, the performance of the trained ML model is little affected up to the 10% variation for the velocity model. Finally, we trained the neural network model considering the results of sensitivity analysis. The ML model trained by training data including errors shows good performance even for the data with 20% error in the peak amplitude.