Seismic vulnerability assessment in urban areas would in principle require the detailed modeling of each single building and the implementation of complex numerical calculations. This procedure is clearly difficult to apply at urban scale where many buildings have to be considered and therefore it is essential to have simplified, but at the same time reliable, approaches to vulnera bility assessment. Among the proposed strategies, one of the most interesting concerns the applica tion of machine learning algorithms, able to classify buildings according to their vulnerability on the basis of training procedures applied to existing datasets. In this paper machine learning algo rithms will be applied to a dataset which collects and catalogs the structural characteristics of a large number of buildings and reports the damage observed in L’Aquila territory during the intense seis mic activity occurred in 2009. A combination of a trained neural network and a random forest algo rithm allows to identify an opportune “ a-posteriori ” vulnerability score, deduced from the observed damage, which is compared to an “ a-priori ” vulnerability one, evaluated taking into account char acteristic indexes for building’s typologies. By means of this comparison an inverse approach to seismic vulnerability assessment, which can be extended to different urban centers, is proposed.