Species- and reactor-aware machine learning for prediction and optimization of microalgal biomass productivity in photobioreactor-based bioresource systems.

Microalgal biomass productivity depends on the complex interactions between the microalgal species, cultivation factors, nutrients, and reactor design. Improving this productivity is important for algae-based bioresource production, biological CO2 utilization, and early-stage screening of photobioreactor operating conditions. This study developed a species- and reactor-aware machine-learning framework to predict and optimize volumetric biomass productivity of Chlorella vulgaris, Spirulina platensis, and Scenedesmus sp. using operational, nutritional, system-level, and reactor-design variables extracted from the literature. Three methods, namely Gaussian Process Regression (GPR), Random Forest (RF) and eXtreme Gradient Boosting (XGB) were trained and optimized using cross-validation and Optuna-based hyperparameter tuning. Among the evaluated methods, GPR achieved the best results on the testing set, R2 = 0.933, RMSE = 0.090 g L⁻1 day⁻1, MAE = 0.056 g L⁻1 day⁻1, and MedAE = 0.031 g L⁻1 day⁻1. Shapley additive explanation (SHAP) values and partial dependence plots (PDPs) highlighted microalgae type, residence time, light intensity, nutrient levels, CO2 concentration, and photobioreactor design parametersas the mostimportant factors influencing productivity. Model-guided optimization suggested species-specific and reactor-constrained cultivation ranges relevant to photobioreactor operation. These optimized ranges were experimentally evaluated under laboratory-constrained photobioreactor conditions for each microalgal species. Obtained experimental productivities for C. vulgaris, S. platensis and Scenedesmus sp. were 0.221, 0.241 and 0.201 g L⁻1 day⁻1, respectively.