Comparison of dry and wet machining in artificial neural network and fuzzy logic

Dry and wet machining each offer distinct trade-offs in surface quality, tool life, cost and environmental impact, and selecting between them is a multi-objective decision that is difficult to model with classical empirical equations. This study experimentally compares dry and wet turning of aluminium alloy Al6061-T6 and develops two complementary predictive models an Artificial Neural Network (ANN) and a Mamdani-type Fuzzy Inference System (FIS) to predict surface roughness (Ra) and material removal rate (MRR) under both conditions. A Taguchi L9 orthogonal array (three factors spindle speed, feed rate, depth of cut each at three levels) was used on a CNC lathe, giving 9 runs per condition (18 observations total: 9 dry, 9 wet); Ra was measured with a Mitutoyo SJ-210 surface tester and MRR was calculated gravimetrically. Both ANN models (dry and wet) used an identical architecture 3 input neurons, 1 hidden layer of 8 neurons, 2 output neurons trained with the Levenberg–Marquardt backpropagation algorithm on min-max normalised data, with a 70:15:15 train/validation/test split and 5-fold cross-validation. The FIS used triangular/trapezoidal membership functions on the same three inputs and a 27-rule Mamdani rule base with centroid defuzzification. The ANN achieved strong predictive accuracy for dry machining (Ra: MAPE 3.6%, R ²  = 0.972; MRR: MAPE 5.9%, R ²  = 0.961) and somewhat lower but still strong accuracy for wet machining (Ra: MAPE 6.2%, R ²  = 0.951; MRR: MAPE 6.7%, R ²  = 0.944), with 5-fold cross-validation confirming generalisation (mean test R ²  = 0.966). The FIS was more balanced across conditions, in particular for MRR (MAPE ≈ 4.5–4.8% versus 5.9–6.7% for the ANN), though it was less accurate than the ANN for Ra prediction. A paired t-test showed a statistically significant reduction in Ra under wet machining relative to dry ( t  = 2.71, df = 8, p  = 0.031; mean improvement 12.4%), whereas the difference in MRR between conditions was not significant ( p  = 0.74), consistent with MRR being calculated gravimetrically from material removed rather than from surface-level thermal effects. Overall, the ANN is recommended for dry-Ra and MRR-critical applications requiring the highest predictive accuracy, while the FIS is recommended where interpretable, rule-based decision support is preferred, particularly for MRR prediction. These findings give manufacturing engineers a validated, quantitative basis for choosing between dry and wet machining strategies and between ANN- and fuzzy-logic-based predictive tools for Al6061 turning.