Modellazione numerica sperimentale a supporto della Artificial Neural Network: valutazione predittiva della resistenza all’accelerazione di collasso tramite curve di progetto degli edifici in muratura

The aim of this research is to develop an innovative and efficient methodology for the expedited assessment of seismic vulnerability in masonry buildings, leveraging the capabilities of artificial neural networks (ANN) integrated with an experimental numerical modeling approach. Following catastrophic seismic events, masonry structures are often significantly compromised, resulting in their classification as unsafe and uninhabitable. Traditionally, the evaluation of such buildings relies on qualitative assessments performed by inspectors, who provide preliminary estimates of structural reliability based on visual inspection and experience. However, this process is inherently subjective and prone to inaccuracies, leading to potential misclassifications that can either overestimate or underestimate the actual risk posed by these structures. To overcome these limitations, the proposed research adopts a machine learning framework, specifically an ANN, to estimate the seismic response of masonry buildings with rectangular geometries. This method allows for a comprehensive and data-driven evaluation of structural vulnerability by incorporating a wide range of building geometries and material properties. The study considers twelve distinct building geometries, twenty-four unique combinations of mechanical parameters, and five different seismic loading directions, resulting in the simulation of 34,560 configurations. These extensive simulations were then summarized through a synthetic polynomial representation, which efficiently encapsulates the complexity of the dataset while enabling streamlined analysis. The ANN was trained, tested, and validated using results from an experimental numerical approach grounded in the Distinct Element Method (DEM), a well-established analytical method for the assessment of structural behavior under seismic loads. The performance of the ANN, when compared to DEM-generated results, demonstrated a high level of accuracy, with predictions differing by approximately 10%. This confirms the viability of using machine learning techniques for the reliable prediction of seismic performance in masonry structures. The primary outcome of this research is the development of a comprehensive database of design curves, which can be employed for the rapid assessment of the seismic vulnerability of masonry buildings. These design curves offer a practical tool for engineers and decision-makers in the aftermath of earthquakes, providing a quantitative and objective basis for classifying buildings as safe or unsafe. The proposed methodology represents a significant advancement over traditional assessment techniques, which are often limited by their reliance on subjective judgment. By combining machine learning with established numerical methods, this research contributes to the development of more reliable and scalable tools for the assessment of building safety in seismicprone areas.