Hybrid Artificial Neural Network-Genetic Algorithm modelling and optimization of multistage palm oil fractional crystallization

Abstract Fractional crystallization of refined, bleached, and deodorized palm oil (RBDPO) is highly sensitive to feedstock variability, particularly fluctuations in iodine value (IV), which influences nucleation, crystal growth, separation efficiency, and olein yield. This study developed and validated response surface methodology (RSM) and hybrid Artificial Neural Network-Genetic Algorithm (ANN-GA) framework for modeling and optimizing multistage palm oil fractional crystallization across RBDPO IV51–IV53. A structured three-stage experimental design encompassing supercooling, nucleation, and crystal growth generated 145 datasets. Stage-specific RSM models quantified process-variable effects and provided an interpretable benchmark, while a multilayer feedforward ANN (7-13-5 architecture) predicted olein yield, olein IV, cloud point (CP), slip melting point (SMP), and stearin IV, achieving R 2 values of 0.79–0.96. GA integration identified feed-specific optimum cooling conditions, predicting olein yields of 83.80%, 83.10%, and 87.98% for RBDPO IV51, IV52, and IV53, respectively. Experimental validation yielded overall prediction accuracies of 91.94%, 96.97%, and 97.75%, with corresponding mean absolute percentage error (MAPE) values of 8.06%, 3.03%, and 2.25%. RSM provided robust interpretability, whereas ANN–GA offered greater flexibility for nonlinear multistage interactions. Collectively, the framework supports feed-responsive optimization of yield and product quality and establishes a basis for adaptive crystallization control and future digitalized palm oil fractionation.