PSI - Issue 79

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

401

a

b

c

Fig. 7. Evaluation metrics of the trained models on training and testing datasets: (a) MLP-1; (b) MLP-2; (c) MLP-3.

During training, the validation loss plateaued at higher values while the training loss continued to decrease, confirming this hypothesis. Given the performance of MLP-1 and MLP-2, it is likely that the global parameters dominate the predictive signal even when both input types are provided. In comparison with existing approaches, MLP-3 adopts an input representation conceptually similar—though not identical—to that of Padhy et al. (2024), who used stacked global descriptors and coordinates from a fixed number of seeding points. Their approach achieves robust predictions by maintaining a constant structural representation, but at the expense of design flexibility. Moreover, their use of seed coordinates instead of nodal coordinates ensures unique structure encoding but small perturbations of the seeds can result in large structural changes. Most importantly, such approaches do not overcome the inherent limitations of MLPs when dealing with variable-size or topology-dependent