VerdaML: an explainable ensemble machine learning framework with AI-driven interactive decision support system for...

Optimisation of microalgal photobioreactor cultivation is essential for the sustainable production of bioproducts, and is challenged by the complex interaction of parameters, low experimental signals and the simultaneous maximisation of biomass productivity and lipid accumulation for biofuel production. This paper introduces VerdaML, an explainable, dual-objective machine learning framework to predict biomass yield of Verrucodesmus verrucosus and model nitrogen-stress induction risk as a proxy for lipid-inductive cultivation conditions as well as provide interaction-based decision support. Domain-informed feature engineering was used to recognise process optimality, stress, and interaction effects using an open-access dataset consisting of 1,080 photobioreactor observations. Biomass yield was predicted using robust regression and ensemble learning, while imbalance-aware classification with SMOTE was used to predict the lipid accumulation potential. The Gradient Boosting regression model yielded a near-ideal yield prediction (R ² = 0.9867, RMSE = 0.1018 g/L, MAE = 0.0763 g/L) and a soft voting ensemble classifier achieved an accuracy of 92.1% (AUC-ROC = 0.9903) for nitrogen-stress risk classification. Five-fold cross-validation confirmed model generalisability, with Gradient Boosting achieving a mean cross-validated R ² of 0.9842 (SD = 0.0024), consistent with the held-out test performance. Joint analysis revealed some trade-offs between fast biomass growth and lipid biosynthesis, and emphasized the drawbacks of single-objective optimization approaches. As a data-driven interactive AI framework, VerdaML makes it possible to optimise photobioreactor processes risk-consciously in adherence to the sustainable bioproduction principles recommended by the United Nations Sustainable Development Goals 7 and 9.