Machine learning-driven multi-objective optimization of Dunaliella salina cultivation for enhanced biomass and β-carotene production.

This study presented an interpretable multi-objective machine learning (ML) framework to navigate the trade-off between biomass accumulation and β-carotene production in Dunaliella salina. Models were developed from 1,494 data points spanning 637 Latin Hypercube Sampling (LHS)-designed regimes, covering eight input variables: temperature, light intensity, salinity, NaHCO3, NaNO3, K2HPO4, putrescine, and cultivation time, with dry cell weight (DCW) and β-carotene yield as output variables. A systematic evaluation of four algorithms, including random forest (RF), extreme gradient boosting (XGBoost), gradient boosting decision tree (GBDT), and artificial neural network (ANN), identified ANN and GBDT as the optimal single-target predictors for DCW and β-carotene yield, respectively. Building on this, their multi-objective versions were developed. The multi-objective ANN, as a unified framework, demonstrated the best predictive performance, achieving an overall test R2 of 0.9758 and accuracy comparable to the specialized single-objective models. Integrated with particle swarm optimization (PSO), the framework generated tailored cultivation strategies (Pareto-optimal and weight-based solutions), which were experimentally validated with all relative errors below 6.67%. The Pareto-optimized strategy enhanced biomass and β-carotene yield by 63.46% and 63.11%, respectively, compared to a non-ML-optimized control. Shapley Additive Explanations (SHAP) analysis revealed cultivation time, salinity, and light intensity to be the most influential factors for model predictions. This work establishes a robust, data-driven paradigm for the intelligent and sustainable optimization of microalgal bioprocesses.