Decoding orthogonal geochemical fingerprints in complex soil matrices via multi-model consensus machine learning for...

Area of Science:

• Forensic Geochemistry
• Chemometrics
• Machine Learning Applications

Background:

• Forensic soil provenance requires scientific objectivity for trial-centered judicial systems.
• High-throughput analysis generates complex data, challenging the extraction of reliable "electronic fingerprints."
• Current chemometric methods suffer from multi-collinear redundancy and "algorithmic dependency trap" biases.

Purpose of the Study:

• To introduce a two-stage feature distillation paradigm to overcome limitations in forensic soil analysis.
• To eliminate algorithmic artifacts and extract resilient, mechanism-driven geochemical fingerprints.
• To develop a legally defensible and reproducible method for soil evidence valuation.

Main Methods:

• Implemented a two-stage feature distillation framework: coarse space sparsification (Least Absolute Shrinkage and Selection Operator) and fine cross-model distillation.
• Utilized a heterogeneous benchmarking suite of nine machine learning algorithms to evaluate feature saliency convergence.
• Focused on identifying algorithmic invariance markers by filtering out mathematical biases.

Main Results:

• Successfully decoupled the soil matrix into a high-dimensional orthogonal geochemical fingerprint space.
• Identified titanium dioxide (TiO2) and silver (Ag) as consensus markers for soil provenance.
• Achieved near-perfect classification fidelity (AUC of 0.998) in cross-validation, demonstrating robustness.

Conclusions:

• The developed framework provides a low-cost, legally defensible, and mechanistically-grounded paradigm for soil evidence.
• TiO2 serves as a geological clock reflecting lithology and weathering, while Ag tracks metallogenic and anthropogenic imprints.
• This approach transitions from empirical modeling to robust feature distillation, meeting stringent reproducibility demands.