Predictive modelling and optimization of material removal rate in WEDM of carbon fiber reinforced polymer composites

Wire electrical discharge machining (WEDM) is one of the few processes capable of cutting carbon fibre reinforced polymer (CFRP) laminates without direct mechanical contact, but predicting the material removal rate (MRR) remains difficult because removal proceeds through a combination of thermal and spark-erosion mechanisms in a highly heterogeneous, poorly conducting workpiece. This study reports an L18 Taguchi experimental campaign on a hand-laid woven-carbon/epoxy laminate and uses it to examine a question that is frequently overlooked in the WEDM modelling literature: how much model complexity a small orthogonal-array dataset can actually support. Five parameters were varied - pulse-on time (T ON ), pulse-off time (T OFF ), spark gap voltage (SV), wire feed (WF) and wire tension (WT). Analysis of variance on the 18 runs shows that MRR is governed by TON (35.6% of the total sum of squares, p = 0.006) and TOFF (34.8%, p = 0.018). Wire feed contributes 8.6% ( p = 0.24), while spark gap voltage (0.4%, p = 0.92) and wire tension (0.3%, p = 0.94) are not statistically significant over the ranges tested. A hierarchy of regression models was then fitted to the same data and assessed by leave-one-out cross-validation (LOOCV) rather than by training fit alone. A three-parameter model in T ON and T OFF gives the best predictive performance (LOOCV R 2 = 0.51), whereas expanding the model to 12 or more coefficients raises the training R 2 above 0.86 while driving LOOCV R 2 negative - that is, the elaborate models predict a new observation less accurately than the sample mean does. A saturated 18-parameter model reproduces the data exactly (R 2 = 1.000) yet has a LOOCV R 2 of -59.1. The practical conclusion is that for an 18-run WEDM dataset, reporting high-order polynomial or multi-term models on the basis of training R 2 is not defensible, and model selection should be based on out-of-sample criteria. The dominance of T ON and T OFF is consistent with discharge energy per pulse and duty cycle controlling removal. The study is limited to a single laminate configuration and a three replicate per run; surface-integrity characterisation of the machined edges is identified as necessary further work and is not claimed here.