PSI - Issue 79

Marco Piacentini et al. / Procedia Structural Integrity 79 (2026) 394–403

400

Fig. 6. Scatter plot of properties (rows) with respect to pairs of generation parameters (columns), colored to help visualization.

3.2. Deep Learning

Among the three models described in Section 2.3, MLP-1 achieved the best overall performance, with R 2 = 0 . 766 for sti ff ness and R 2 = 0 . 882 for strength on the test dataset (Fig. 7a). This confirms that the global generation parameters contain su ffi cient information to capture first-order dependencies of the e ff ective mechanical properties. In contrast, MLP-2, which uses only the node coordinates, performs substantially worse ( R 2 = 0 . 414 for sti ff ness and R 2 = 0 . 559 for strength; Fig. 7b). This degradation does not imply that the coordinate data are uninformative, but rather likely reflects the inability of fully connected MLPs to e ff ectively process a list of nodal coordinates of largely incoherent size, which also causes incoherent numbering. Furthermore, connectivity information is also missing. As a result, the MLP lacks tool to build even approximately robust mapping from the inputs. The hybrid MLP-3, which combines both input types, achieved slightly lower test performance than MLP-1 despite reaching better training results (Fig. 7c). This behavior indicates overfitting: the larger representational capacity—resulting from the increased number of neurons—allows better fitting of the training data without improving generalization to unseen structures.