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

The global transition toward trial-centered judicial systems demands unprecedented scientific objectivity and verifiability in forensic soil provenance. While emerging high-throughput analytical methodologies generate rich multi-modal datasets, extracting robust "electronic fingerprints" from these complex matrices remains a formidable challenge. Current chemometric applications are heavily hindered by severe multi-collinear redundancy and, more critically, the "algorithmic dependency trap", wherein the identified critical markers are heavily biased by the specific mathematical architecture of a chosen machine learning classifier. To overcome these limitations, this study introduces a two-stage feature distillation paradigm designed to eliminate algorithmic artifacts and extract resilient, mechanism-driven fingerprints. In the initial phase of coarse space sparsification, a Least Absolute Shrinkage and Selection Operator regularization scheme is executed under strict linear constraints to suppress multi-collinear noise and isolate a highly discriminative candidate subset. In the subsequent phase of fine cross-model distillation, this subset is exposed to a heterogeneous benchmarking suite of nine cutting-edge algorithms. By evaluating cross-model feature saliency convergence, we filter out mathematical biases and capture true algorithmic invariance markers. Remarkably, this distillation framework successfully decouples the convoluted soil matrix into a high-dimensional orthogonal geochemical fingerprint space governed by two consensus markers: titanium dioxide (TiO₂) and silver (Ag). Geochemical interpretation confirms that the conservative, high-field-strength TiO₂ functions as an immutable geological clock reflecting parental lithology and chemical weathering baselines, whereas the chalcophile element Ag acts as a sensitive tracker of localized metallogenic backgrounds and superimposed anthropogenic imprints. This orthogonal system achieves near-perfect classification fidelity with an Area Under the Curve (AUC) of 0.998 in cross-validation. By transitioning from empirical predictive modeling to robust feature distillation, this framework provides a low-cost, legally defensible, and mechanistically-grounded paradigm for soil evidence valuation that directly satisfies the stringent reproducibility demands of modern jurisprudence.