Sviluppo di metodi innovativi di machine learning per la previsione di terremoti in Italia e Messico

Responsabili di progetto

Luciano Telesca, Wen Yu

Accordo

MESSICO - CINVESTAV - Centro de Investigación y de Estudios Avanzados del Instituto Politécnico Nacional

Bando

CNR/CINVESTAV biennio 2019-2020 2019-2020

Dipartimento

Scienze del sistema terra e tecnologie per l'ambiente

Area tematica

Scienze del sistema Terra e tecnologie per l'ambiente

Stato del progetto

Nuovo

Proposta di ricerca

For decades the idea that earthquakes may be forecasted was based on the recurrence interval for characteristic earthquakes, that repeat almost periodically; this idea was abandoned, when predicted events did not occur. Since then, several forecasting methods were proposed, based on measuring earthquake-related signals (geoelectrical, radon, ionospherice,..), or on parameters derived by statistical analysis of seismicity (b-value, epicentral fractal dimension, etc.). Seismic precursors were observed in laboratory faults (Goebeletal., 2013) and recognized retrospectively before the occurrence of large earthquakes (Bouchonetal., 2016). The plethora of forecasting methods just indicates how complex earthquakes are and how still challenging the problem of prediction is. In 2011 the International Commission on Earthquake Forecasting for Civil Protection (ICEFCP) concluded that there is "considerable room for methodological improvements in this type of (precursor-based failure forecasting) research", indicating very clearly how crucial is the development of innovative methodologies for forecasting earthquakes. The always improving of seismic networks has permitted to get a huge amount of data, and thus has renewed hope that earthquake forecasting could be a feasible task, if combined with the improving of forecasting methods, as requested by the ICEFCP.
Very recently, Rouet-Leduc et al. (2017) proposed a machine learning technique, the random forest, to forecast earthquakes generated in laboratory faults, by modeling the relationship between input signal (features derived from acoustic emission) and corresponding output label (the time remaining before the next failure event, derived from shear stress). Random forest represents one of the machine learning algorithms that, however, need to be adjusted when they return an inaccurate prediction.
An advancement of machine learning is represented by the "deep learning", which is capable of determining on its own if the predictions are accurate or not, due to: 1) the much larger number of hidden layers than the classical neural networks, 2) its capability to "learn" something intrinsic about the data without the help of a target or label vector, 3) its capability to train the model in the sense of probability. Such capability of the deep learning model to self-adjust is really important in the context of seismic forecasting, since the complexity of the earthquake process has so far led all the classical existing prediction models to fail. The manner in which deep learning works is not so different from what the other machine learning algorithms do, like random forest; the only difference is that they do the work much better and efficiently.
Practically, using existing earthquake data (defined in terms of epicentre, time of occurrence and magnitude) one or several machine learning models (random forest and deep learning models) will be trained. Then these models will be used to predict parameters of the seismicity, linked with the occurrence of large earthquakes. So, the idea is to use existing earthquakes to predict large earthquakes.
Since machine learning (and in particular random forest and deep learning) is a general model, in the context of the present project, we will study how to specify or modify this model such that it can work well for the earthquake prediction, how and why do some modifications work, how accurate the predictions are. To do it, we will combine machine learning with a moving window technique to construct time-continuous series of several well-defined seismic parameters (like the coefficient of variation, the b-value of the Gutenberg-Richter law, the fractal dimension of epicentres, the scaling exponent of the Allan Factor, etc.), and use this time series to perform the predictions.
The performance of the machine learning (random forest and deep learning) earthquake forecasting models will be tested in Italy and Mexico, for which earthquake forecasting represents a research priority, due to the great socio-economic impact the occurrence of potentially damaging earthquakes could have in territories highly densely populated (like Mexico) or with numerous historical buildings/cities that even a medium intensity earthquake could irreversibly damage or destroy (like Italy).
The results of the project will have important societal impact: the availability of robust machine learning forecasting tools could integrate the traditional seismic hazard tools that are mainly based on estimating the Gutenberg-Richter parameters and completeness magnitude. They could help civil protection actions contributing to select those areas where more necessary are risk reduction interventions and improving emergency preparedness. The patterns selected by applying such methods as the more effective, could be proposed for a forward testing in the next years and in areas more prone to large events in the next future in both countries. This project will have also a great scientific impact on seismological modeling, because the assessment of more appropriate space-time models of seismic processes will prompt their usage in many other seismic areas worldwide.
This project entails a joint research effort among experts from the Mexican team, which has an outstanding experience in machine learning, confirmed by the leading of several important projects and by papers published in well reputed journals, and from the Italian team, which complements with a longstanding activity in seismic analysis and skill in advanced time series analysis. Such synergy is fundamental to attain the aimed project goals and to envisage a cooperation for future project submissions. The outcomes of the project (papers, knowledge transfer, staff mobility, exchange of methods and human resources) will open new avenues in the Mexican-Italian cooperation in the context of seismic forecasting.

Obiettivi della ricerca

1) to develop and apply machine learning forecasting methods to Italy and Mexico, in relationship with the forecasting of large earthquakes; 2) to improve the understanding of space-time dynamics of seismic process in Italy and Mexico by using innovative time series based forecasting methods; 3) to evaluate likelihood of claimed relationships between space-time patterns of the parameters calculated by using machine learning forecasting methods and impending earthquakes; 4) to produce forecasting maps in Italy and Mexico on the base of the application of innovative forecasting tools; 5) to foster young researchers and Ph.D students of both involved institutions in the project performance through the interaction of the members of both teams; 6) to disseminate intermediate and final results through papers submitted to peer-reviewed journals with IF and conferences and by a dedicated internet website in order to give the maximal visibility to the project.

Ultimo aggiornamento: 06/06/2025