Deep learning forecasting of large induced earthquakes via precursory signals

www.nature.com/scientificreports OPEN Deep learning forecasting of large induced earthquakes via precursory signals Vincenzo Convertito 1, Fabio Giampaolo 2, Ortensia Amoroso 3 & Francesco Piccialli 2* Precursory phenomena to earthquakes have always attracted researchers’ attention. Among the most investigated precursors, foreshocks play a key role. However, their prompt identification with respect to background seismicity still remains an issue. The task is worsened when dealing with low-magnitude earthquakes. Despite that, seismology and, in particular real-time seismology, can nowadays benefit from the use of Artificial Intelligence (AI) to face the challenge of effective precursory signals discrimination. Here, we propose a deep learning method named PreD-Net (precursor detection network) to address precursory signal identification of induced earthquakes. PreD-Net has been trained on data related to three different induced seismicity areas, namely The Geysers, located in California, USA, Cooper Basin, Australia, Hengill in Iceland. The network shows a suitable model generalization, providing considerable results on samples that were not used during the network training phase of all the sites. Tests on related samples of induced large events, with the addition of data collected from the Basel catalogue, Switzerland, assess the possibility of building a real-time warning strategy to be used to avoid adverse consequences during field operations. Deterministic earthquake prediction is still far from being possible due to the complexity and the limited knowledge of the system geoscientists have to deal with. However, considerable steps have been made toward the identification of reliable precursory phenomena that can allow us to understand if the system is evolving toward a critical/unstable state. Although their proper identification is still debated, foreshocks have been referred to as the most obvious premonitory phenomenon preceding e arthquakes1 thus representing the most promising candidate2,3. Foreshocks have been interpreted as the failing of populations of small patches of fault as they reach a critical stress that eventually but not necessarily become large e arthquakes4 or as a part of the nucleation process which ultimately leads to the mainshock5,6. Distinguishing precursors, such as foreshocks from ordinary seismic activities, as for example earthquake swarms or switching, is not trivial and may hamper their usefulness in reliable earthquake p rediction3,7,8. In practice, identification of the precursory phase of large earthquakes is mainly based on the analysis of earthquake catalogues and more recently from the analysis of geodetic data, and from waveform similarity a nalysis9. As an example, timely and accurate earthquake location can help to envisage earthquakes space migration. Besides, seismic catalogues of tectonic earthquakes allow investigating statistical features that characterise foreshocks with respect to mainshocks and aftershocks. For instance, changes in the slope of the Gutenberg-Richter r elation10, namely the b-value, or power-law time-to-failure fitting have been analyzed as discrimination t ools5,11. Whatever the adopted tool, all the implemented approaches require empirical criteria for selecting time and space windows and seismicity occurrence models, such as time-dependent Poisson12 or ETAS model13,14, to identify groups of earthquakes as candidates for being classified as foreshocks. Nevertheless, the significant amount of data collected in the last years and the advances in computer hardware have given Artificial Intelligence (AI) a great popularity in almost all scientific areas, including seismology. AI techniques can learn significant patterns from the data to generate models to support and sustain human expertise. In particular, different approaches from Machine Learning (ML) and Deep Learning (DL) are successfully applied in the study of earthquakes15–18 and their d etection19–21, with also some application in the field of induced seismicity22, and particular aimed at trying to anticipate the location of the areas where earthquakes are expected to o ccur23. According t o24 earthquake catalogues collected by using AI have reached an unprecedented quality and detail that can help seismologists formulate and test new hypotheses about precursors of large earthquakes. 1 Istituto Nazionale di Geofisica e Vulcanologia, Osservatorio Vesuviano, Naples, Italy. 2Department of Mathematics and Applications “R. Caccioppoli”, Univeristy of Naples Federico II, Naples, Italy. 3Department of Physics “E.R. Caianiello”, University of Salerno, Fisciano, SA, Italy. *email: Scientific Reports | (2024) 14:2964 | https://doi.org/10.1038/s41598-024-52935-2 1 Vol.:(0123456789) www.nature.com/scientificreports/ In this study, we propose a strategy for the identification of precursors of the large induced earthquake in a sequence based on a supervised DL approach that can be implemented in real-time applications. It should be noted that the use of induced s eismicity25 brings with it the intrinsic difficulty that earthquakes may not occur as foreshock/mainshock/aftershock sequences but rather may occur as sequences of earthquakes with magnitudes close to each other. However, the choice to analyse induced seismicity was guided by the fact that seismic catalogues of induced earthquakes in the magnitude range of interest for the present study are characterised by a larger number of events and a lower minimum magnitude of completeness compared to tectonic seismicity. This is a key point since, at least for tectonic earthquakes, seismic catalogue incompleteness can produce artefacts in observed rates of both foreshocks and aftershocks14 making the statistical approach ineffective. In addition, the time and space evolution of the seismicity can be correlated to known field operations. We investigated a set of specific features, which are generally considered prognostic of the earthquake preparatory phase (i.e. the minimum magnitude of completeness Mc, the b-value, moment magnitude ( MW ), the moment rate, duration of events’ group, the coefficient of variation CoV , the Fractal Dimension, the NearestNeighbour distance, the Shannon’s Information Entropy and associated uncertainties). We analysed data collected in three regions of induced earthquakes, namely The Geysers (TG) geothermal fi eld26,27, Cooper Basin (CB) geothermal reservoir28,29, and the Hengill geothermal field (HG)30. Moreover, to further assess the generalization capabilities of the approach, it has been tested on a stand-alone series, that is, not used for training and validation, extracted from the Basel (BS) catalogue, related to injection operations in 2006. The proposed DL approach relies on a three-step system: data labelling; a Neural Network (NN), namely the PreD-Net, for classification; and finally, a warning strategy. In the data labelling step, a label is associated with (...truncated)