MACHINE-LEARNING (NEURAL NETWORK–GENETIC ALGORITHM) OPTIMIZATION OF THE SIMULTANEOUS NADES–MICROWAVE–ULTRASOUND EXTRACTION OF PECTIC AND PHENOLIC FRACTIONS FROM CACAO (Theobroma cacao L.) POD HUSK

Cacao (Theobroma cacao L.) pod husk (CPH), the main residue of cocoa processing, is rich in pectin and phenolic compounds, but their recovery usually relies on long, solvent-intensive extractions that target one fraction at a time. Natural deep eutectic solvents (NADES) combined with microwave and ultrasound offer a green and rapid alternative, while machine learning can model non-linear, multi-factor processes. The objective was to optimize the simultaneous recovery of pectic and phenolic fractions from CPH (CCN-51) by NADES-based microwave–ultrasound-assisted extraction using an artificial neural network (ANN) and a genetic algorithm (GA). Among nine NADES screened, choline chloride:glycerol gave the highest pectic extract yield (73.0 ± 4.1%, dry basis) and choline chloride:malic acid the highest total phenolic content (TPC; 43.2 ± 1.6 g GAE/100 g dry flour), whereas choline chloride:oxalic acid offered the best balance (62.7 ± 2.8% and 40.5 ± 0.2 g GAE/100 g) and was selected. Six factors (water:NADES and solvent-to-solid ratios, microwave power and time, ultrasound time and temperature) were studied in a 2⁽⁶⁻²⁾ fractional factorial design with three center points (19 treatments, n = 3). The precipitated pectic extract was weighed and the TPC of the phenolic fraction was measured by Folin–Ciocalteu (dry-flour basis). Because the gravimetric extract retains NADES, results are reported as pectic extract yield, not pectin yield. Pectic extract yield (27.8–82.6%) increased with microwave time and decreased with water:NADES ratio and microwave power (p < 0.05), while TPC (22.3–37.0 g GAE/100 g) increased with solvent-to-solid ratio (p < 0.01). A linear main-effects model showed significant lack of fit (p < 0.0001), indicating factor interactions. A multilayer perceptron (6-4-3-2, ensemble of five L2-regularized networks) trained on all 57 observations fitted both responses (R² = 0.99 and 0.88); however, leave-one-condition-out cross-validation (17 folds) showed limited predictive capacity for untested conditions, as expected for a small two-level dataset. A real-coded GA maximizing Derringer's desirability within the equipment's operating levels identified a water:NADES ratio of 3:1 (v/v), 40 mL/g, 350 W, 56 s of microwave and 22 min of ultrasound at 40 °C, predicting 83.1% pectic extract yield and 37.0 g GAE/100 g. This optimum is a model prediction supported by its proximity to the best experimental treatment (82.6 ± 1.2% and 36.8 ± 0.2 g GAE/100 g); its experimental verification is in progress. Short processing conditions (≈1 min of microwave and 22 min of ultrasound) were thus identified for the simultaneous recovery of both fractions. Pending determination of galacturonic acid content, degree of esterification, purity and functional properties, the pectic extract represents a potential future ingredient for active pectin-based films and coatings.

Funding: This work was funded by PROCIENCIA (CONCYTEC, Peru), Contract N° PE501095554-2025-PROCIENCIA.