In this study, 1,238 stream‑sediment samples from the Pariz and Chahargonbad areas of southwestern Kerman Province were analyzed to identify geochemical anomalies. The dataset was enhanced using principal component analysis (PCA), independent component analysis (ICA) and polynomial feature expansions, followed by normalization to capture both linear and non‑linear variability. Four anomaly–background separation models were evaluated: Isolation Forest, DBSCAN, Autoencoder and Variational Autoencoder (VAE). Model hyper‑parameters (e.g. the number of layers and neurons per layer) were tuned via Bayesian optimization in Python using the Optuna library. During tuning, the scoring metrics were reconstruction error for the Autoencoder; likelihood‑based reconstruction error for the VAE; the anomaly score for Isolation Forest; and the mean distance of each sample to the cluster core for DBSCAN. For spatial presentation, samples were assigned to upstream drainage basins, and each sample’s result was propagated to its upstream area; the outcomes were then compared with known mineralization points across the region. Systematic hyper‑parameter tuning improved the detection of geochemical signatures and highlighted zones with mineral potential. Comparative analysis showed that robust feature engineering combined with Bayesian optimization significantly enhances anomaly‑detection accuracy. These findings underscore the practical value of advanced machine‑learning techniques in geochemical exploration and provide a framework for more targeted field investigations in mineral‑rich regions.
 
Introduction
Efficient recognition of geochemical anomalies underpins mineral exploration and environmental assessment. Rising data dimensionality and nonlinearity challenge classical uni/ multivariate thresholds and geostatistics, motivating machine-learning (ML) and deep-learning (DL) approaches that capture complex, multi-element signals and integrate ancillary data [1]. Autoencoder-family models have proven effective for weak, multivariate anomalies using reconstruction-error criteria, and adversarial variants further enrich latent representations. Yet performance hinges on feature engineering and hyperparameter selection; prior work often relies on defaults or ad-hoc tuning. This study targets those gaps by: (i) expanding geochemical features (PCA, ICA, polynomial terms), (ii) optimizing IF, DBSCAN, AE and VAE via Bayesian search, and (iii) validating anomaly maps against known mineralization in a well-endowed metallogenic belt (Sarcheshmeh–Meydouk district).
 
Methodology and Approaches
Geological setting and data: The area lies within the Urumieh–Dokhtar magmatic arc, with Oligocene–Miocene calc-alkaline intrusions hosting world-class porphyry Cu deposits (e.g., Sarcheshmeh, Meydouk). A dataset of 1,238 stream-sediment samples was compiled for multi-element analysis.
Preprocessing and feature engineering: Outliers were screened (interquartile-range criteria), element values log-transformed and min-max normalized. Correlation structure and hierarchical clustering guided feature design. Five PCA and five ICA components were appended, and second-order polynomial terms were generated for Cu-correlated elements to capture nonlinear interactions
 
Models and objective functions
- Isolation Forest (IF): tuned (estimators, max_ samples, contamination) to maximize average decision-function score (higher = more normal), sharpening separation from anomalies
- DBSCAN: tuned (ε, min_samples, k) under a composite objective balancing outlier fraction and neighborhood-distance compactness to form geologically coherent clusters while isolating true outliers.
- AE (deep): encoder–decoder with monotonically contracting hidden sizes; Optuna minimized a composite of mean reconstruction error and its standard deviation to favor accurate and stable reconstructions
-VAE: latent Gaussian prior (μ, σ) trained with MSE reconstruction + β·KL divergence; layers, hidden width, latent size, batch size, and output activation were optimized for minimum validation loss
Bayesian optimization (Optuna) adaptively explored hyperparameter spaces, improving efficiency over grid/random search and providing consistent, data-specific configurations. Model scores were aggregated to catchment outlets (SCB) to generate anomaly probability classes for mapping and validation against 39 known mineralized points.
 
Results and Discussion
Optimized configurations and comparative performance
Bayesian tuning yielded clear performance gains across all detectors.
- Isolation Forest: high-anomaly area 21%, capturing 10/39 occurrences.
- DBSCAN: 47% area, 24/39 occurrences.
- Autoencoder (AE): 36% area, 24/39 occurrences -  the best balance of precision (compact footprint) and recall of known mineralization.
- Variational Autoencoder (VAE): 42% area, 24/39 occurrences.
Spatially, the AE geochemical anomaly map outlines coherent belts coincident with known porphyry systems and prospective intrusive centers. Feature augmentation (PCA+ICA+polynomial terms) consistently lifted signal-to-noise and stabilized training, especially for deep models.
Model-specific behavior.
 IF favored compact delineations but under-captured weak, diffuse halos. DBSCAN effectively separated dense background clusters from dispersed outliers but inflated high-anomaly areal extent at tuned ε/min_samples. AE/VAE reduced over-generalization while preserving mineralization hits; AE slightly outperformed VAE in footprint control under the chosen β and architecture
 
Conclusions
Robust feature engineering coupled with Bayesian hyperparameter optimization substantially improves unsupervised geochemical anomaly detection. 2) Among tested models, the Autoencoder provides the most favorable precision–recall trade-off for the Pariz–Chahargunbad dataset, capturing 24/39 known occurrences within a 36% high-anomaly footprint 3) VAE and DBSCAN achieve comparable recall but at larger footprints (42–47%), whereas Isolation Forest is conservative but misses a portion of known targets. 4) SCB-based spatial propagation yields interpretable maps aligned with drainage-mediated signal transport.