This study critically evaluates the predictive capability of artificial neural network (ANN), polynomial regression, support vector regression (SVR), and random forest (RF) models for estimating the mechanical properties of recycled polypropylene/spent coffee grounds (r-PP:SCG) composites using only SCG content and silane concentration as input variables. The dataset consisted of 80 observations obtained from 16 formulation conditions, and repeated grouped 5-fold cross-validation was applied to reduce replicate dependency, data leakage, and overly optimistic performance estimation. Polynomial regression was used as a conventional statistical baseline, while SVR and RF were employed as machine-learning benchmark models. The ANN model showed limited-to-moderate and property-dependent predictive performance, with R² values of 0.438, 0.475, -0.079, 0.259, and 0.375 for tensile strength, Young’s modulus, elongation at break, impact strength, and hardness, respectively. Comparative analysis showed that ANN did not consistently outperform the baseline and benchmark models, confirming that increased model complexity does not necessarily improve prediction under small, homogeneous, and composition-only datasets. The poor predictability of deformation and energy absorption-related properties indicates the importance of microstructural descriptors such as filler dispersion, porosity, density variation, and interfacial adhesion. Therefore, ANN should be regarded as a preliminary screening and diagnostic tool for sustainable composite formulation rather than a definitive design-optimization model without external validation and additional structural descriptors.
Nithikarnjanatharn et al. (Mon,) studied this question.