PSI - Issue 78

Alina Elena Eva et al. / Procedia Structural Integrity 78 (2026) 387–394

393

Fig. 6. Damage to the facade

Fig. 7. Transfer learning with DANN

ing approach was applied, with the masonry represented as an equivalent homogeneous material (Tab. 1). Since the properties of model are based on theoretical assumptions and not intended to capture the detailed interaction between bricks, mortar, and the resulting composite behavior of masonry, consistency between individual material phases was not the primary modeling objective. Therefore, in the micromodeled region, the mortar properties were approximated using the global mechanical characteristics of the tested masonry from the Bologna facade, ensuring a reasonable match at the macroscopic level while simplifying the calibration process for this first approach. The numerical model also includes additional structural components such as the wooden lintel and two floor slabs, modeled as distributed non-structural masses to capture dynamic e ff ects without altering the overall sti ff ness of the structure. The modeling approach is preliminary, and the induced damage in the full-scale facade model is minimal(see Fig. 6). The damage is observed on the panel, likely due to the uncalibrated masonry properties in the facade model. When considering strain variations relative to the healthy facade condition, healthy deformation patterns between panel and facade are well aligned (see Fig. 7). Although the neural network was able to detect the presence of damage in the fac¸ade, the damage severity was too low to be classified with su ffi cient accuracy. This suggests that enriching the dataset with ad ditional damage classes—spanning a wider range of severity levels—would improve the model’s ability to distinguish between subtle and more pronounced damage states.