Artificial intelligence based optimization of electrical discharge machining parameters for minimizing surface roughness of aluminium 6061 boron carbide metal matrix composites

Modern manufacturing increasingly demands the optimization of non-linear responses during the Electrical Discharge Machining (EDM) of Metal Matrix Composites (MMCs). This study investigates the surface integrity of an A6061 alloy reinforced with 6 wt% B 4 C (Boron Carbide), synthesized via the stir casting route. Seventy-four experiments were designed using a Central Composite Design (CCD) framework employing a copper electrode and hydrocarbon dielectric oil, and seven controllable machine settings were varied: discharge current (I), electrode working time (P on ), electrode lift time (P off ), pulse-on time (T on ), gap voltage (G v ), duty factor (τ) and flushing pressure (F p ). P on and P off are servo and flushing cycle timers expressed in seconds, and are physically distinct from the discharge duration T on , which is expressed in microseconds. Analysis of variance established that only three factors are statistically significant at the 95% confidence level: P on (30.22% contribution), T on (19.82%) and I (19.59%), together accounting for 69.63% of the total variation. The quadratic response surface model explained 91.08% of the variation in surface roughness (SR). A feed-forward back-propagation Artificial Neural Network with a 7–10–1 architecture, tansig hidden activation, purelin output activation and Levenberg–Marquardt training achieved a correlation coefficient of R = 0.9975 and a mean squared error of 0.0477 µm 2 , reducing the root mean squared error from 0.3022 μm (RSM) to 0.2184 μm, an improvement of 27.7%. A Genetic Algorithm (GA) and Simulated Annealing (SA) were additionally applied to the RSM polynomial in order to benchmark surrogate-driven global search; both converged to boundary solutions and are therefore reported as comparative benchmarks rather than as the recommended optimizer. The measured surface roughness ranged from 2.2 μm at 3 A and 50 µs to 5.9 μm at 15 A and 250 µs. At the optimum combination of 3 A current, 2 s electrode working time, 1 s electrode lift time, 50 µs pulse-on time, 40 V gap voltage, 90% duty factor and 0.2 kgf/cm 2 flushing pressure, the ANN predicted SR = 2.1 μm against a confirmed experimental value of 2.2 μm, a relative error of 4.5% compared with 12.7% for RSM, 12.0% for GA and 34.8% for SA, a 64.6% reduction in prediction error relative to RSM. Scanning electron microscopy of the optimally machined surface revealed shallow overlapping craters, a thin discontinuous recast layer, few resolidified globules and no micro-cracking. The proposed intelligent model provides a robust framework for enhancing machining precision in abrasive MMCs.