Mechanism-infused machine learning for scattered metal recovery: SHAP decodes multiscale physicochemical drivers in refractory matrices

Abstract

The inefficient recycling of electronic waste exacerbates critical metal shortages and environmental risks. Chlorination roasting, a common method for recovering scattered metals, faces challenges due to fluctuating material compositions and complex process variables, rendering traditional optimisation approaches costly and inefficient. This study developed a machine learning predictive model to enhance recovery efficiency for scattered metals. Based on 18 physicochemical and process parameters, the Extreme Gradient Boosting model achieved an R² of 0.97 on the test dataset. SHAP analysis revealed the causal mechanisms linking features to recovery rates. Experiments utilised fly ash as an endogenous chlorine source for systematic optimisation of germanium (Ge) recovery, achieving a leaching rate of 97.2%. Predictions for indium (In) validated the model's accuracy; it also demonstrated robust generalisation capability for gallium (Ga) – absent from training data – with a relative prediction error of 6.7%, while achieving 100% substitution of toxic chlorinating agents in all cases. This research provides a universal data-driven solution for the efficient, environmentally sound recovery of multiple scattered metals.