Learning source, path and site effects: CNN-based on-site intensity prediction for earthquake early warning

To provide timely and accurate seismic alerts for potential users during the earthquake early warning (EEW) process, several algorithms have been proposed and implemented. Some of the most common rely on the characterization of the earthquake magnitude and location, and then use a ground motion model to forecast shaking intensity at a user's location. It has been noted that with this approach the scatter in the forecasted intensities can be significant and may affect the reliability and usefulness of the warnings. To ameliorate this, we propose a single station machine learning (ML) algorithm. We build a four-layer convolutional neural network (CNN), named it CONIP (Convolutional neural network ONsite Intensity Prediction), and test it using two data sets to study the feasibility of seismic intensity forecasting from only the first few seconds of a waveform. With only limited waveforms, mainly P waves, our CONIP model will forecast the on-site seismic intensity. We find that compared with existing methods, the forecasted seismic intensities are much more accurate. To understand the nature of this improvement we carry out a residual decomposition and quantify to what degree the ML model learns site, regional path, and source information during the training. We find that source and site effects are easily learned by the algorithm. Path effects, on the other hand, can be learned but will depend largely on the number, location, and coverage of stations. Overall, the ML model performance is a substantial improvement over traditional approaches. Our results are currently only applicable for small and moderate intensities but, we argue, could in future work be supplemented by simulations to supplement the training data sets at higher intensities. We believe that ML algorithms will play a dominant role in the next generation of EEW systems.