Machine learning optimization between pullulan recovery and adsorptive removal of melanin using date seed biochar

The co-production of melanin during Aureobasidium pullulans fermentation compromises the purity and functionality of pullulan. To address this, acid-activated date seed biochar was evaluated for the selective removal of melanin from crude pullulan. A central composite design of experiment was adopted to vary solution pH, contact time (CT), and biochar concentration (BC) in the batch adsorption process, while quantifying melanin and pullulan recovery through a rapid, nondestructive chemometric Partial Least Square Regression (PLSR) model (R2CV 0.99) built on designed blends. Hyperparameter-optimized Artificial Neural Network (ANN) models trained on Gaussian noise-augmented data demonstrated excellent prediction performance for both melanin adsorption (R2: 0.97, R2CV: 0.95) and polymer loss (R2: 0.99, R2CV: 0.99). Explainable machine learning tools, such as Shapley Additive Explanation (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), were used to interpret the developed ANN models. Genetic algorithm-based multi-objective optimization suggests melanin removal of 87.56% and polymer loss of 14.55% under optimal biosorption conditions (i.e. pH: 5.16, CT: 36.59 min, and BC: 20.00 g L−1), closely matching the observed melanin removal (86.85%), polymer loss (17.42%) in the validation run. Combining green adsorbents, chemometric tools, and a machine learning model provides a robust, scalable solution for downstream processing of biopolymers.