Improving geochemical anomaly recognition associated with mineralization via counterfactual causal inference

Identifying geochemical anomalies in lithologically complex regions remains challenging due to the superposition of multiple geological processes, which complicates the separation of regional background from mineralization-related signals. Here, we propose a counterfactual-inspired framework that reframes an anomaly as the deviation of an observed value from its expected background under a non-mineralization scenario, thereby extending conventional unsupervised outlier detection approaches by incorporating spatial prior information from known mineral occurrences. This framework integrates three stages: (1) median quantile regression forest (QRF-M) to isolate a clean background, (2) causal forest to quantify spatial association with known mineral occurrences, and (3) extraction of counterfactual anomalies by predicting background concentrations exclusively from non-mineralized areas. The method yielded three main findings. First, QRF-M suppresses outliers, producing unbiased regional geochemical background estimates, and identifies Fe 2 O 3 , SiO 2 , and Ti as primary background controls. Second, the average treatment effect on the treated (ATT = 12.08 ppm) exceeds the average treatment effect (ATE = 3.02 ppm), suggesting spatially heterogeneous enrichment near known deposits. Third, the counterfactual method achieves an anomaly recognition rate of 73.7 % within 5 km of known occurrences—higher than conventional threshold (47.7 %) and QRF-M (50.5 %) methods—while reducing false anomalies and highlighting three prospective areas not flagged by comparison methods. However, because the same known occurrences are used both to define the treatment variable and to evaluate spatial coincidence, this efficiency index is subject to dual-use evaluation bias and should be regarded as a retrospective brownfield benchmark for well-explored brownfield settings, not as an independent prospective validation or evidence of universal algorithmic superiority. Its primary value lies in detecting weak mineralization signals in low-background areas that conventional methods often miss, though the highlighted targets require field verification.