A Machine-Learning Approach for the Reconstruction of Ground-Shaking Fields in Real Time

ABSTRACT Real-time seismic monitoring is of primary importance for rapid and targeted emergency operations after potentially destructive earthquakes. A key aspect in determining the impact an earthquake reconstruction ground-shaking field, usually expressed as ground-motion parameter. Traditional algorithms compute field from punctual data at stations relying on prediction equations computed estimates location magnitude when instrumental are missing. The results such then subordinate to evaluation magnitude, which can take several minutes. To fill temporal gap between arrival estimate these parameters, a new data-driven algorithm that exploits information station only introduced. This algorithm, consisting ensemble convolutional neural networks (CNNs) trained database maps produced with traditional algorithms, provide their associated uncertainties real time. Because CNNs cannot handle sparse data, Voronoi tessellation selected peak ground parameter recorded used input CNNs; site effects network geometry accounted using (normalized) VS30 map map, respectively. developed method robust noise, changes over time without need retraining, resolve multiple simultaneous events. Although having lower resolution, obtained statistically compatible ones methods. fully operational version running servers Department Mathematics Geosciences University Trieste, showing real-time capabilities handling Italian strong-motion outputting resolution 0.05° × 0.05°.