Development of machine learning approaches for real-time ground-shaking maps reconstruction

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 of an earthquake is the reconstruction of the ground-shaking field, usually expressed as the ground motion parameter. Traditional algorithms compute the ground-shaking fields from the punctual data at the stations relying on ground-motion prediction equations (GMPEs) computed on estimates of the earthquake location and magnitude when the instrumental data are missing. The results of such algorithms are then subordinate to the evaluation of location and magnitude, which can take several minutes. The integration of neural networks in the processing workflow has been assessed as a viable way to speed up ground-shaking map reconstructions. New data-driven algorithms that leverage the information recorded at the stations have been developed to provide real-time estimates of the ground-shaking fields and fill the temporal gap between the arrival of the data and the estimate of these parameters. The first developed algorithm, called ShakeRec, is an end-to-end model consisting of an ensemble of convolutional neural networks (CNNs) trained on a database of ground-shaking maps produced with traditional algorithms. It provides estimates of the ground-shaking maps and their associated uncertainties in real time. Site effects are accounted for using a normalized VS30 map as input. A second algorithm called ShakeRec-hybrid has been developed to integrate neural networks in the ShakeMap® workflow to speed up the current ground-shaking maps evaluation process while overcoming some of the intrinsic limitations of ShakeRec. The core idea is to adopt the multivariate normal distribution (MVN) method used in ShakeMap® for the IM interpolation and use a neural network, based on the convolutional conditional neural process and replacing the GMPEs, to estimate the IM conditional expected value and uncertainty at the target sites based only on data available in real-time. Considering a correlation function independent of the epicentral distance and the event magnitude, the whole workflow became independent from the evaluation of the source parameters resulting in a considerable speed-up. By reusing the ShakeMap® framework, the complexity of the model is reduced with improvements in the interpretability of the results. To further improve the interpretability of the results, the local site effects have been accounted for using amplification factors. The developed methods have been tested for their reconstruction capabilities (using real data), robustness to noise, network geometry changes over time, and considering multiple simultaneous events. The adaptability of ShakeRec-hybrid to different areas is mainly linked to the availability of large datasets to train the network. To overcome this possible limitation two different approaches have been considered. The first approach consists of generating synthetic data to train the neural network: this provides an efficient way to obtain a large training dataset with a high level of control over all the desired parameters, allowing also to target specific scenarios and develop a more balanced dataset. The second approach consists of fine-tuning a model developed for an area to a different one where only a small dataset of recorded data is available, showing promising results. The two approaches could also be combined either by pre-training the neural network on a synthetic dataset and with a final fine-tuning stage on real data or by developing larger and more balanced datasets mixing real and synthetic data. The proposed algorithms have been implemented and integrated into BRTT Antelope showing real-time capabilities in handling the data flow from the integrated Italian strong motion network stations and a web service has been developed to disseminate the results in real-time.