Bio-Inspired Optimization of AWJM Parameters for Enhanced Machining of Glass and Carbon Fiber-Reinforced Composites

Abstract

Abrasive water jet machining (AWJM) of fiber-reinforced composites requires balanced control of surface roughness, kerf width, and material removal rate. This study proposes a statistically bound, benchmarked, and reproducible regression-desirability optimization workflow for the glass fiber-reinforced composite (GFRC) and carbon fiber-reinforced composite (CFRC) using the published Taguchi L9 AWJM data set of Rao et al. as a secondary-data benchmark. Abrasive mass flow rate (AMFR), cutting speed rate (SR), and stand-off distance (SOD) were modeled against surface roughness (Ra), kerf width (Kw), and material removal rate (MRR) using first-order ordinary least-squares regression. Model adequacy was evaluated using coefficient inference, ANOVA, residual diagnostics, error metrics, and leave-one-out cross-validation. The fitted surrogates were coupled with desirability aggregation and optimized using the Whale Optimization Algorithm (WOA), followed by benchmarking against GA, PSO, and NSGA-II and bootstrap uncertainty analysis. MRR showed the strongest statistical structure. For GFRC, AMFR significantly influenced MRR (β = 21.887, p = 0.009; R2 = 0.801), while for CFRC, SR was significant for MRR (β = 8.216, p = 0.040; R2 = 0.697). In contrast, Ra and Kw showed weak linear predictability, with low R2 and negative LOOCV R2 values. Within the bounded OLS surrogate–desirability formulation, WOA identified local compromise settings of 95 g/min, 360 mm/min, and 2.5 mm for GFRC (D = 0.6052), and 95 g/min, 397.45 mm/min, and 4.5 mm for CFRC (D = 0.6310). GA and PSO reached identical optima, while NSGA-II gave closely comparable solutions, indicating optimizer consistency rather than WOA superiority. The framework supports bounded AWJM parameter selection under small-sample constraints. However, because Ra and Kw showed weak out-of-sample predictive validity, their optimized values are reported only as local, low-confidence OLS surrogate estimates; reliable Ra/Kw optimization requires nonlinear or expanded experimental modeling.