GNGTS 2016 - Atti del 35° Convegno Nazionale

52 GNGTS 2016 S essione A matrice Some reasoning on the improvement of the ETAS modeling at the occurrence of the 2016 Central Italy sequence A.M. Lombardi Istituto Nazionale di Geofisica e Vulcanologia, Roma, Italy Introduction. The use of time-dependent models has been introduced recently in Italy to produces forecast seismicity maps (Marzocchi et al., 2012, 2014). This type of activity is expected by operational earthquake forecasting (OEF), a set of procedures for gathering and disseminating information about the time dependence of seismic hazards (Marzocchi et al. , 2014). The models that participate to the short-term OEF in Italy (Marzocchi et al. , 2014), are rather well established and mainly focused on the short-term spatio-temporal clustering of earthquakes. Therefore, their great skill is the capacity to describe the evolution of sequences, whereas they have a poor predictive ability before the occurrence of the main events. The establishment of robust short-term OEF procedures requests to increase our understanding of how and with how much precision earthquakes are predictable. A multi- disciplinary scientific approach involving seismological, geodetic, geological and historical techniques can significantly improve knowledge of the cause of earthquakes and strengthen the physical basis for earthquake predictability. Case studies may serve for the improvement of the OEF procedures, which should be able to forecast the spatio-temporal evolution of sequences and to improve the knowledge about the occurrence of main events. Firstly, I present an analysis of the first part of the 2016 central Italy sequence by mean of the Epidemic Type Aftershock Sequences (ETAS) model (Ogata, 1998; Lombardi, 2015, 2016a, 2016b), to check the performance of the model on very provisional data. Particular care is taken in a sensitivity analysis of forecasts to both the input parameters of the model and to the occurrence history. Second, I provide some preliminary points of interest for a possible improvement of this type of models. The analysis of the 2016 central Italy sequence. On August 24, 2016, at 01:36 UTC, a magnitude ML 6.0 earthquake hit part of the central Italy area, close to the villages of Amatrice and Accumuli, one of the most hazardous regions of Italy. The largest aftershock, of magnitude ML5.4 (the only one event with magnitude above 5.0, for now), occurred near the Norcia city, about 1 hour after the main event. Here, I apply the ETAS model to study the evolution of the first part of the seismic sequence. The model is set on the seismicity data of the official INGV bulletin (www.iside.rm.ingv.it) , fromApril 16, 2005 up toAugust 24, 2016, occurred in the region [12.7-13.8E, 42.0-43.5N] and with magnitude ML ≥ 2.5, including the whole 2009 L’Aquila sequence. This region contains the smaller area [12.95-13.5E, 42.45-43.0N], interested by the 2016 central Italy sequence. The version of the ETAS model used here has been implemented in SEDAv1.0 (Statistical Earthquake Data Analysis), a statistical software freely provided via the Zenodo open access platform (https://zenodo.org/record/55277; Lombardi, 2016a, 2016b). To check the performance of the model, firstly, I compare the observed number of events with what expected by 10000 ETAS simulated catalogs (number of events test, Lombardi, 2016a). The test is performed for N D overlapping interval times of one day {D i , i=1,...,N D }, updated each hour, starting from the time occurrence of the mainshock (August 24, 2016, 01:36:32, ML6.0) up to September 15, 2016. This test also allows measuring the sensitivity of the model to the occurrence history, which fully participates to ETAS calculations. Fig. 1 shows the comparison between the expected and the observed number of events with magnitude above MF=2.5, 3.0 and 4.0. It shows the median expected number of events (black line) together with the 95% confidence bounds (red lines), quantifying the uncertainty of predictions. The results show a strong underestimation in the first 8 hours, due to data incompleteness. Then the observations are above the median forecasts, but inside the confidence