Constrained Multicriteria Optimization of Injection-Molded Biocomposites: A Comparison of Genetic Algorithm and Bayesian Optimization

Introduction: Using response models calibrated from injection-molded polypropylene (PP)/alfa-fiber composites containing maleic-anhydride-grafted PP (MAPP), this study compares a steady-state genetic algorithm (GA) and constrained Bayesian optimization (BO) for lightweight reusable logistics panels.

Methods: Density was minimized over alfa-fiber content, MAPP content, and effective retained fiber length. The latter was treated as a post-processing material parameter indirectly influenced by compounding and molding, not as a directly settable control variable. Micromechanical and empirical relations predicted density, fiber volume fraction, Young’s modulus, tensile strength, and elongation at break. Screening feasibility required (E \geq 5.50) GPa, (\sigma_u \geq 38.00) MPa, (\varepsilon_b \geq 1.50%), and (V_f \leq 0.40). Both methods used the same 32 initial designs and 45 sequential additions, giving 77 evaluated designs per method. Results represent one reproducible stochastic realization per algorithm using fixed random seeds.

Results: BO’s best feasible design contained 35.75 wt% alfa fiber, 0.01 wt% MAPP, and (L_f=4.947) mm; predicted responses were (\rho=1.0317) g cm(^{-3}), (E=5.502) GPa, (\sigma_u=50.98) MPa, (\varepsilon_b=2.40%), and (V_f=0.263). GA found 36.97 wt% alfa fiber, 2.33 wt% MAPP, and (L_f=5.000) mm; corresponding responses were 1.0372 g cm(^{-3}), 5.710 GPa, 51.98 MPa, 2.35%, and 0.274. Both designs satisfied all constraints. In this realization, BO’s predicted density was 0.53% lower than GA’s and its stiffness was closer to the active 5.50 GPa limit. The near-zero MAPP result is model-dependent because the Pukánszky interaction parameter was MAPP-independent; it is not a validated processing dosage.

Conclusions: Under the matched design budget, BO found the lighter feasible design in this single realization. Repeated-seed statistics, MAPP-dependent interface modeling, and experimental validation are required before general algorithmic or manufacturing conclusions.