Revealing Key Mechanisms in Multimechanism Interplay for ReO₄^- Removal: A Knowledge-Data Dual-Driven Framework for...

Efficient removal of ⁹⁹TcO₄^- from nuclear waste is critical for radioactive waste disposal, yet current adsorbents face limitations in capacity and mechanistic clarity due to multimechanism interference. Herein, we propose a knowledge-data dual-driven (DKD) machine learning (ML) framework to elucidate the dominant mechanism for ReO₄^- (a nonradioactive surrogate for ⁹⁹TcO₄^-) adsorption on covalent organic frameworks (COFs). By integrating domain-knowledge (expressed as mathematical descriptors for five adsorption mechanisms) into the ML model, the DKD approach achieved higher predictive accuracy (R² = 0.93) and interpretability than the purely data-driven model (R² = 0.91). SHAP analysis of the DKD model quantitatively identified the electrostatic interaction as the primary mechanism, contributing 66.7% to ReO₄^- uptake. Guided by this insight, we first broke through the conventional building block charge enhancement by directly introducing charge at the linkage, designing Tb-APDC-M─an imine-linked COF with an ultrahigh charge density. It achieved a record ReO₄^- adsorption capacity of 1689.78 mg g^-1 (pH = 7, T = 298.15 K, dosage = 0.5 g/L), significantly exceeding the previous maximum of 1262 mg g^-1. Spectral and DFT calculations confirm that the exceptional performance stems from the high charge density imparted by the iminium linkage. This work highlights the DKD framework's efficiency in pinpointing key mechanisms and enabling targeted adsorbent design for ⁹⁹TcO₄^- removal.