Robust source tracing for solid waste via machine learning-enabled mineralogical fingerprinting

As global urbanization and consumption surge, the mismanagement of solid waste has emerged as a critical bottleneck for sustainable development, preventing the recovery of valuable resources and threatening ecological security. Tackling this issue requires precise source tracing, yet conventional methods are often ineffective due to process fluctuations and complex waste mixing. Herein, inspired by genetic mineralogy, we propose “Anthropogenic Typomorphic Mineral Assemblages (ATMA)” as a novel diagnostic tracer, systematically validating its effectiveness in the case of heavy-metal hazardous waste. Integrating data covering 159 waste codes from 25 sources, we employed a knowledge-guided noise-injection strategy to select robust mineralogical fingerprints. The resulting machine-learning model demonstrated high accuracy (> 0.9) under noisy test conditions while effectively mitigating overfitting. SHAP analysis effectively revealed ATMAs for each source, confirming their existence and consistency with industrial metabolic patterns. This work provides an interpretable, noise-resilient mineralogical forensics method for waste tracing, and uncovers the value of ATMA as records of material metabolism in the technosphere.