PSI - Issue 41

11

America Califano et al. / Procedia Structural Integrity 41 (2022) 145–157 Author name / Structural Integrity Procedia 00 (2019) 000–000

155

Figure 8 - ROC curves for the test attempts carried out on data from plane z = 8 (a) and plane z = 2 (b). The numbers for the tested configurations are reported in bold near the curves. The legend describes the values of AUC computed for each ROC curve. The FNR vs. FPR curves are reported for the twelve configurations for the test attempts carried out on plane z = 8 (c) and plane z = 2 (d). The legend described the values of the area under each of these new curves, namely 1-AUC. 5. Conclusions In this work, a preliminary attempt in implementing machine learning algorithms to predict the development of climate-induced damage in already-cracked panel paintings has been proposed. A parametric 3D FE model of a panel painting made by a wooden support and a layer of gesso with craquelures has been implemented; several numerical simulations have been carried out varying the geometrical parameters of the model, to assess the possible development of further damage conditions through the Strain Energy Density failure criterion. Based on preliminary geometrical considerations, the problem has been simplified and the attention has been focused on 2D models extracted from the simulated 3D configurations. Extracted data have been used to train and test powerful machine learning algorithms, known as eXtreme Gradient Boosting (XGBoost) machines, on predicting the arising of safe or unsafe conditions by simple classification tasks. Several tests have been implemented in order to evaluate the robustness and the sensitivity of the trained XGBoost classification model and, by evaluating commonly adopted scores, promising preliminary results have been obtained. Future scenarios will deal with the optimization of the XGBoost model (by fine-tuning its hyperparameters), the comparison with other approaches (deep neural networks, for example), and the implementation