GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 2.1 GNGTS 2024 Some thoughts on the testing phase of seismic hazard models W. Marzocchi University of Naples Federico II The intrinsic scientific nature of any probabilistic seismic hazard analysis (PSHA) – including national seismic hazard model (NSHM) – implies that its credibility has to be based (only) on a rigorous and extensive testing phase. However, this phase presents many challenges of different nature and limitations. Here we discuss in detail and with some real examples some of the most important ones. The long-term time scale of NSHM (50 years) limits the possibility to validate a model, i.e., to check if NSHM describes satisfactorily independent data that were never used to build the model. However, past observations can be used to check the consistency of NSHM, i.e., if the model is able to explain the past observations. The difference between validation and consistency is not only semantic, because building and testing a model with the same past data can easily lead to overfitting, which may boost improperly the credibility of a model; conversely, overfitting is ruled out when testing model with independent data (validation). Being the consistency tests much more common, it must be kept in mind that the outcomes of this testing phase may not reflect the real goodness of a model, because of the unavoidable and often unquantifiable overfit. Owing to the limited number of recorded ground shaking observations, the consistency of NSHM is also checked by analysing partial outcomes of the model, which may have more data available for testing (the number of data is linked to the power of the test). Specifically, a NSHM is a complex model with two major components: an earthquake rupture forecast model (ERFM) and a ground motion model (GMM). Hence, the consistency of NSHM can be also checked analysing its ERFM capability to describe satisfactorily the space-time distribution of the past large earthquakes. It goes without saying that this is a sine qua non condition, i.e., a reliable NSHM model has to be composed by a ERFM that describes satisfactorily the past earthquake occurrences, but the opposite is not true: an ERFM that describes well the past large earthquake occurrences does not necessarily lead to a good NSHM if the GMM is wrong. In essence, testing the consistency of ERFM can rise a red flag on the credibility of a NSHM model, but it cannot guarantee for its reliability. Too often (and unfortunately), NSHM is still based on declustered earthquake catalogs, mimicking the ground shaking of the so-called mainshocks. However, a physics-based