SEISMICITY-CONSTRAINED MULTIMODAL MULTITASK MACHINE LEARNING FOR GEOLOGICAL FAULT CHARACTERIZATION

Geological fault detection and characterization plays a key role in assessing and early warning seismic hazards. The spatiotemporal distribution of anthropogenic or natural microseismicity (MEQ) are observed to be highly correlated with geological fault locations. We develop a novel geological fault detection and characterization machine learning (ML) model based on a multimodal, multitask neural network (NN) that integrates both seismic migration image and MEQ location information. In specific, our seismicity-constrained fault characterization model has two encoder branches and one multitask decoder branch. One encoder branch encodes a subsurface reflectivity image generated by seismic migration, while the other encoder branch encodes a MEQ location image. We represent each MEQ by a multidimensional Gaussian function with a finite spatial support. The decoder branch integrates the learned features from both the seismic migration image encoder and the MEQ location image encoder, and eventually estimate multiple fault attributes, including the spatial location, strike, and dip angle of each image pixel/voxel, in an end-to-end manner. We validate this novel ML model, and the results show that the integration of MEQ information can improve the accuracy and fidelity of automatic geological fault characterization compared with conventional models that are based solely on seismic migration images. The NN may therefore serve as a powerful tool in identifying and characterizing large-scale fluid injection and hidden geological faults.