POST-EARTHQUAKE BUILDING PERFORMANCE: MACHINE LEARNING FOR 2016–2017 CENTRAL ITALY EARTHQUAKES

In post-earthquake situations, the rapid evaluation of the buildings' performances and usability may play a crucial role, especially for an effective emergency management. This research discusses a machinelearning-based workflow to cope with the observed post-earthquake buildings' usability data concerning the 2016–2017 central Italy earthquake sequence. Specifically, the data-driven model is expected to provide an useful supporting tool for predicting the building's usability immediately after an earthquake event. In order to improve the predictive capabilities of the model, the intensity measures have been encompassed in the postearthquake usability data coming from the rapid evaluation AeDES forms related to the Abruzzi region (Italy). To improve the performances of the intelligent classifiers, a binary classification problem has been formalized to distinguish between no-damaged or slightly damaged structures, i.e. which can be almost immediately occupied again, from those that are partially or entirely unusable, which, on the other hand, require significant restoration interventions or complete demolition. The adopted methodologies for an effective data treatment due to the imbalanced nature of the dataset have been discussed in order to attempt to provide fair training of the models accounting for the minority class. The current study highlighted the main still existing challenges in dealing with a strongly imbalanced dataset likewise the one under investigation. Nevertheless, the promising results emphasized that a properly calibrated machine learning model may provide a useful tool to support the decision-making process in emergency conditions estimating the building performances and usability in postevent seismic scenarios