Interpretable Machine Learning for Elucidating Component Synergy in a Solid Waste-Derived Multicomponent Adsorbent for Heavy Metal Removal.

Waste magnesia-carbon bricks and drinking water treatment residues are used by mechanochemical method to prepare a multi-component adsorbent (MC-WT) composed of LDHs, graphite, and siliceous particles to remove As(V), Cr(VI), Pb(II), and Cu(II) from soil. Adsorption equilibrated within 180 min and followed pseudo-second-order kinetics and the Generalized-Langmuir isotherm model. Site energy distribution indicated average site energies of 14.9-26.7 kJ/mol for MC-WT and 9.3-15.2 kJ/mol for the raw wastes, revealing stronger affinity for the metals. However, the removal contribution of each coexisting component was difficult to identify, hindering the rational design of the raw material ratio. With the component contents as input variables, random forest provided the highest accuracy (R2 > 0.99), and Shapley additive explanation with characterization resolved the contribution of each component: LDHs contributed most through interlayered ion exchange and chemical precipitation, graphite through weak redox and surface complexation, and siliceous particles least through weak electrostatic interaction. Partial dependence analysis identified optimal LDHs and graphite contents of 3.39-3.48 and 1.92-2.03 g. Removal was affected by pH and coexisting ions, and column leaching, fractionation, and regeneration confirmed the retention of heavy metals in soil, providing a basis for the component-level design of waste-based adsorbents.