From physics-based simulation to ground motion models using Machine-Learning Estimator for Ground Shaking Map

The south of the Iberian Peninsula, particularly the Baetic System, is one of the most seismically active regions of the Iberian Peninsula. Its complex seismotectonic configuration causes recurrent moderate to strong earthquakes, posing a significant hazard to society and the built environment, requiring rapid and accurate post-event assessment of ground-motion intensity.  These high-risk areas coincide with densely populated areas of Murcia, such as Lorca, or the province of Almeria. In addition, population dynamics vary significantly between summer and winter, due to seasonal tourism and residential tourism, which increases vulnerability and the need for rapid and accurate assessments following an earthquake. To address this need, the Machine Learning Estimator for Ground Shaking Maps (MLESmap) was developed as a rapid-response framework that combines high-quality physics-based simulations with Machine Learning techniques to infer spatially distributed ground-motion intensity measures within seconds after earthquake initiation. Trained on a large ensemble of synthetic seismic scenarios, MLESmap provides near real-time predictions of ground-motion intensity fields, such as acceleration levels and shaking patterns.

Our methodology incorporates both offline and online phases in a comprehensive workflow. It begins with the generation of a synthetic training data set generated by the CyberShake platform. Then  predictor characteristics are extracted before the validation and learning stages. The result is a model that can be used for fast inference validated with start-of-art methodologies and available real data  .

To evaluate the influence of surface representation on model performance, synthetic simulations are carried out using both 1D and 3D seismic velocity models, allowing for a systematic comparison of their impact on training and prediction accuracy. In addition, different learning strategies are explored, as for example, multi-objective approaches that allow for the simultaneous estimation of multiple measures of ground motion intensity. These analyses quantify the influence of velocity model dimensionality and training strategy on the performance of MLESmap predictions

Funded by the European Union. This work has received funding from the European High Performance Computing Joint Undertaking (JU) and Spain, Italy, Iceland, Germany, Norway, France, Finland and Croatia under grant agreement No 101093038, ChEESE-2P, project PCI2022-134980-2 funded by MCIN/AEI/10.13039/501100011033 and by the European Union NextGenerationEU/PRTR)