A Two-Stage Residual Modeling Framework for Data-Driven Prediction of Uranium Adsorption Performance in Aqueous Environments.

Given the severe threat of uranium contamination, the application of theoretical and experimental models to develop efficient adsorbents for radioactive wastewater treatment is in strong demand. Herein, the intrinsic characteristics of materials (e.g., crystalline structure, surface area and zeta potential, etc.) and key parameters dominating uranium adsorption process (e.g., dosage, solution pH, etc.) were systematically identified. Nine machine learning algorithms were optimized using Bayesian optimization and cross-validation. To capture prediction biases and uncover latent patterns, a well-trained Random Forest was proposed as the residual optimization learner, forming a two-stage optimization (TSO) process. Results showed that the Multilayer Perceptron (MLP) achieved statistically significant improvement, with a 33.7% increase in R2 and 44.4% reduction in RMSE. To quantify the contribution of the critical features in the decision-making process and explore their potential influence patterns and marginal effects on the predicted output, SHAP and partial dependence plot (PDP) analyses was coupled to interpret the residual model. This framework enables adaptive correction of prediction biases across different algorithms, offering a new algorithmic support and methodological foundation for the intelligent design and performance optimization of novel materials for providing data-driven reference for rapid assessment of uranium adsorption behavior under laboratory conditions.