Explainable Deep Learning for Real-Time Prediction of Uniform Hazard Spectral Acceleration for On-Site Earthquake Early Warning

Summary Earthquake early warning systems are designed to provide critical seconds of warning before strong ground shaking, facilitating emergency mitigation efforts. Existing methods, such as neural networks and ground motion prediction equation-based approaches, rely on manually defined parameters and physics-based computations, which introduce human bias and hinder the efficiency of real-time applications. Furthermore, current studies primarily focus on scalar metrics such as peak ground acceleration and peak ground velocity to evaluate earthquake impacts. These metrics are limited to measuring ground shaking intensity and fail to capture the spectral characteristics of ground motion. Therefore, a ground-motion and structural-oriented deep learning-based model is proposed to predict uniform hazard spectral acceleration values across 111 periods ranging from 0.01 to 20 seconds. The framework is initially trained and evaluated on 17,500 ground-motion records from the crustal Next Generation Attenuation West 2 project. Spectral acceleration values are predicted by two subsets: deep learning-based uniform hazard spectral acceleration models 1 and 2. These models effectively utilize feature information from the initial seconds of seismic waveforms, eliminating the need for empirically defined parameters. Two deep learning-based models are developed for two datasets representing two distinct broad geographical regions. Both models utilize a similar deep-learning architecture but vary in input settings and hyperparameters to account for regional seismic characteristics. To assess the model's goodness-of-fit between observed and predicted values, as well as its generalization ability, we rigorously compare the two models with the latest data from the U.S. Geological Survey Earthquake Hazard Toolbox and the Japanese Strong-Motion Earthquake Network, respectively. An explainable artificial intelligence technique has been applied to better understand the framework and analyze how individual input features influence the outputs of the trained models. Integrating cutting-edge deep learning technologies into ground motion and engineering seismology reveals the significant potential of the model in enhancing real-time early warning systems. This integration also provides valuable support to various end-users involved in seismic monitoring, facilitating well-informed decisions in both real-time and near-real-time scenarios.