Machine learning assisted prediction and process validation of electrochemically induced phosphorus recovery from wastewater

The disruption in the natural phosphorus cycle has enhanced the problem of depletion of phosphorus rocks and its overabundance in the wastewater, leading to various environmental problems. Electrochemical phosphorus recovery from wastewater has gained attention due to the production of high-purity phosphorus precipitates that can be utilized to relieve the pressure on natural sources. However, the comprehension of process parameters using modern machine learning algorithms are needed to maximize the phosphate recovery efficiency before scale-up. Hence, this study aims to predict electrochemical phosphorus recovery using different machine learning (ML) models: Linear regression, Lasso regression, Ridge regression, Adaboost, eXtreme Gradient Boost (XGB), Random Forest, and Support Vector Regression. The dataset for 16 input parameters considering the wastewater, reactor, and reaction characteristics from the literature was used to envisage its influence on electrochemical phosphorus recovery. The XGB proved to be the robust ML model as it outperformed other models in testing with R-2 (0.98) and RMSE (33.54). The influence of the input parameters was studied using the feature importance, dependence and summary plot. Current density, pH, inter-electrode distance, electrolysis time and initial phosphorus concentration were some of the most influential input parameters affecting the phosphorous recovery. The process validation shows the R-2 of 0.78 between the predicted values from the model and the experimental result. Overall, this study could help to predict electrochemical phosphorus recovery from the wastewater to facilitate optimization, upscaling, and commercialization of the technology to close the phosphorus loop.