Machine learning and nutrient circular economy recovery for microalgae Chlorella Vulgaris biomass production

Capturing and reusing critical nutrients from wastewater streams has gathered international interest in combating eutrophication, and offers a vehicle for greater insulation against volatile price shocks on fertilisers and conservation of finite resources. This research paper explores the use of machine learning (ML) models for predicting nutrient removal (i.e., recovery through nutrient assimilation into biomass) effectiveness specific to the microalgae species Chlorella Vulgaris . Several inputs such as the initial Total Phosphorous (TP), initial Total Nitrogen (TN), pH, biomass concentration, days exposure, light intensity, air flow rates, CO 2 concentration in flow streams, temperature and working volumes were modelled as inputs against predicting for TP, TN removal and biomass production accuracies across multiple linear regression (MLR), multilayer perception (MLP), random forest (RF) regressor and eXtreme Gradient Boosted (XGBoost) Decision Trees (DT) ML models. The models were trained with 4472 datapoints collected from nine literature papers and the performance were evaluated. The simulations show that biomass recovery predictions using XGBoost was the best with an R 2 , RMSE and MAE of 0.98, 0.138 and 0.076, respectively. Followed by RF regressor for biomass recovery predictions with an R 2 , RMSE and MAE of 0.961, 0.186 and 0.099, respectively. For TN and TP removal efficiencies, SHapley Additive exPlanations (SHAP) analysis shows that the inputs with the greatest sway in final prediction accuracies were found to be days exposure, initial TP, biomass and pH. Meanwhile, biomass significant input variables were air flow temperature, CO 2 concentration in airflow influent, days for growth and pH. The study shows that ML models can be used to predict biomass nutrient recoveries without significantly involving carbon as a critical dataset component for an ML nutrient circular economy (CE). • XGBoost biomass yielded superior prediction performance over comparison models. • RF performed the second best for biomass production prediction. • Multiple linear regression neural network had lowest performance. • Feature important parameters for total nitrogen, total phosphorous and biomass were identified. • ML provides strong models for predicting nutrient, microalgae recovery for CE.