Combined machine learning - 3D physics based approach for building damage evaluation: the case of L'Aquila 2009.

This work combines artificial intelligence tools and physics-based simulations with the aim to develop a hybrid tool to support pre- and post-earthquake emergency phases. As a case study, we examine the 6.1 Mw earthquake that struck and partially destroyed the city of L’Aquila and many surrounding villages on 6 April 2009. Numerical simulations were performed using SPEED (SPectral Elements in Elastodynamics with Discontinuous Galerkin), which is a spectral element based software, for earthquake simulation on local and regional scale. The simulated values of peak ground velocity and peak ground acceleration, together with those provided by ShakeMaps, are included in a dataset containing information on 3.060 buildings damaged during the 2009 earthquake. Three different damage distributions were considered, obtained from the six damage levels assigned in the original dataset, ranging from D0 (light or no damage) to D5 (heavy damage or collapse). The results show that the peak ground velocity and peak ground acceleration obtained from numerical simulations have a significantly greater impact on the level of damage recorded than those derived from ShakeMaps. In fact, simulated intensity measures are among the four most important features for all tests considered, while variables obtained from ShakeMaps do not even appear in the top twelve, consequently, they may not serve as a reliable proxy for building damage assessment. The simulated intensity measures have been used to assess the building damage reaching a maximum accuracy of the 83% on a balanced dataset.