Infer-GANomaly: A Generative AI Framework for Seismic Event Discrimination via Physics-Informed Hybrid Attention

Abstract

Distinguishing between natural and artificial seismic events, referred to as seismic event discrimination, is pivotal in geohazard monitoring, but accurate discrimination remains technically challenging. To address this issue, in this letter, we propose Infer-GANomaly, a novel, high-performing seismic event discrimination framework that leverages a generative deep learning (DL) technique with physical information about seismic events. In our approach, we strategically recast the seismic event discrimination problem as an anomaly detection task and adopt the cutting-edge GANomaly pipeline as backbone. The key breakthrough in our method lies in the enhancement of the generator, which reconstructs seismic representations in the GANomaly backbone, by developing and deploying a novel, innovative hybrid attention module that incorporates the so-called physics-informed conditioning mechanism and linear modulation. Another innovation lies in the design of a custom training loss function that integrates the physical information for effective adversary learning of the proposed Infer-GANomaly. Experimental results on the STEAD dataset demonstrate the superiority and efficacy of our method over other competing approaches.