Multi-objective Optimization of Green MnO₂–Napier Fibre/Epoxy Nanocomposites Using an Integrated BBD–RSM–ANN–MOGWO–TOPSIS Framework
Abstract
This study presents an integrated statistical and machine-learning framework for optimizing the mechanical performance of epoxy composites reinforced with alkali-treated Napier grass fibres and green-synthesized MnO₂ nanoparticles. A three-factor Box–Behnken design comprising 15 experimental runs was used to investigate molding pressure (2–6 MPa), molding temperature (80–120 °C), and MnO₂ loading (0.5–1.5 wt.%). The experimental responses ranged from 53.4 to 73.0 MPa for open-hole tensile strength (OHTS), 3.42 to 4.65 GPa for elastic modulus (EM), and 0.0161 to 0.0225 mm/mm for failure strain (FS). The developed quadratic models were highly significant (p < 0.0001), exhibited R² values above 0.997, and showed nonsignificant lack of fit. RSM desirability optimization identified a condition of 4.73 MPa, 104.5 °C, and 1.23 wt.% MnO₂, predicting an OHTS of 73.79 MPa, EM of 4.78 GPa, and FS of 0.0161 mm/mm. An MLP–ANN trained using the Levenberg–Marquardt algorithm achieved training R² values of 0.9957–0.9970 and leave-one-out cross-validation R² values of 0.9528–0.9595. The validated ANN was integrated with the multi-objective grey wolf optimizer to maximize OHTS and EM while minimizing FS, which was defined as a non-beneficial criterion representing the selected preference for a stiffer, lower-deformation composite. While RSM desirability provided a direct compromise based on the fitted quadratic models, ANN–MOGWO generated 100 non-dominated Pareto solutions, from which TOPSIS selected the final balanced solution using predefined response weights. The selected condition was 5.82 MPa, 103.8 °C, and 1.27 wt.% MnO₂, with predicted OHTS, EM, and FS values of 70.50 MPa, 4.62 GPa, and 0.0161 mm/mm, respectively. Unlike conventional RSM desirability optimization, the integrated approach enabled transparent evaluation and ranking of multiple trade-off solutions, with experimental confirmation errors of 0.97–1.24% for the selected condition. The proposed framework facilitates data-driven exploration of processing conditions and reduces the need for additional optimization experiments beyond the initial 15-run design. The optimized composites show potential for lightweight, moderately loaded structural components where enhanced stiffness, controlled deformation, and sustainable material selection.