AI-Driven Prediction of Uranium Adsorption on Acid-Modified Biochar: Integrating Large Language Models with Interpretable Machine Learning

Efficient uranium recovery from radioactive wastewater is a pressing challenge in sustainable nuclear energy development. Here, we present an innovative AI-driven framework that integrates large language models with machine learning to optimize uranium adsorption by acid-modified biochar. By automatically extracting and structuring literature data, we compiled a high-quality dataset comprising 589 experimental data points. Among four tree-based ensemble models, CatBoost delivered the best performance (test R2 = 0.98, RMSE = 24.78). Feature importance indicates that adsorption conditions (71.51%) as the most influential factors, significantly outweighing biochar preparation conditions (15.27%) and physicochemical properties (13.22%). SHapley Additive exPlanations (SHAP) analysis further provided insights into how key features influence uranium adsorption, identifying important variables and their impact patterns. Finally, we developed a user-friendly graphical user interface that enables rapid, intelligent prediction of uranium adsorption capacity, supporting data-driven experimental design. This study provides a referenceable, AI-powered solution for radioactive wastewater treatment and offers a transferable framework for the intelligent remediation of other environmental pollutants.