Interpretable transfer learning for predicting electrochemical phosphorus recovery from wastewater

Muhammad Waqas Nadeem, Irfan Hassan, Muzammil Hussain, Farhan Bashir, Hock Guan Goh, Mian Umer Shafiq

2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.1016/j.sajce.2026.100858
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Wastewater contains valuable resources such as phosphorus, which can be recovered and reused, thereby reducing the environmental impact of wastewater treatment. Phosphorus recovery from wastewater is essential to decrease eutrophication, protect aquatic ecosystems, and promote sustainable development. Among various recovery methods, electrochemical phosphorus (P) recovery has emerged as a sustainable and efficient approach for resource recovery and pollution control. However, its implementation across diverse operating conditions has yielded heterogeneous, often incomplete datasets, making the accurate prediction of recovery efficiency challenging. This study introduces an interpretable transfer learning (TL) framework based on feature-space transfer from unsupervised representation learning to supervised prediction. It combines a stacked autoencoder (SAE) for feature extraction with an XGBoost regressor for prediction. To address missing values, different imputation methods, including statistical, machine learning, and deep learning, were applied to reconstruct the dataset. The predictive model within the TL framework was then trained and tested on each imputed dataset to evaluate how different imputation strategies influence predictive performance. Among these, the dataset imputed with a recurrent neural network (RNN) yielded the highest accuracy, highlighting the importance of deep learning–based data reconstruction under heterogeneous experimental conditions. Using the same predictive model architecture, the TL framework achieved an R² of 0.99 and an RMSE of 21.06, results that compare favourably with those obtained with other imputation methods (average R² = 0.99 ± 0.01, RMSE = 27.3 ± 4.8). Model interpretation reveals that electrolysis time (2–2880 min), stirring rate, and calcium concentration (0–640 mg L⁻¹) were the most influential factors affecting phosphorus recovery. These findings demonstrate that combining deep learning–based imputation with interpretable transfer learning significantly enhances predictive accuracy and provides practical insights for optimizing electrochemical phosphorus recovery in wastewater treatment systems.