Answers are grounded in this paper and cite their sources.
Natural fiber-reinforced polymer composites (NFRPCs) have emerged as sustainable alternatives to synthetic composites; however, reliable prediction of their mechanical properties remains challenging because of variability associated with fiber architecture, matrix type, and processing conditions. This study presents a property-specific artificial neural network (ANN) framework comprising four independently optimized feedforward ANN models. The models were trained on a large open-access experimental dataset comprising 590 uniaxial tensile specimens representing ten composite configurations, including flax, jute, and carbon fiber systems with unidirectional, cross-ply, woven, and short-fiber architectures in epoxy and vinyl ester matrices. Twelve geometric and stress-based features extracted from the early portion (0–1% strain) of the stress–strain response were used as model inputs. The framework accurately predicted elastic modulus (R 2 = 0.9932), ultimate tensile strength (UTS, R 2 = 0.9969), strain at UTS (R 2 = 0.7480), and modulus of toughness (R 2 = 0.9795). Across all targets, the average test-set R 2 was 0.9294, while the mean 5-fold cross-validation R 2 reached 0.8874. These results compare favorably with previously reported ANN and ensemble ML approaches for fiber-reinforced composites, which have achieved R 2 values in the range of 0.93–0.97, while the present framework additionally provides property-specific SHAP-based interpretability across a substantially larger (n = 590) and more architecturally diverse experimental dataset than prior studies. SHapley Additive exPlanations (SHAP) analysis revealed that stress-based features were the dominant contributors to elastic modulus and UTS predictions, whereas failure strain and fiber architecture primarily governed toughness. The lower prediction accuracy for strain at UTS was consistent with the high experimental scatter observed in cross-ply and short-fiber composites (coefficient of variation: 17–22%), indicating an inherent noise-limited prediction ceiling. The suggested framework offers a useful tool for quicker characterization of natural fiber composites by providing an understandable and effective method for calculating important mechanical properties from partial tensile response data.
No takes yet. Share an insight, caveat, or question.
Tiwari et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: