GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration¶

Conference: CVPR 2026 arXiv: 2603.13068 Code: github.com/yihaoding/geochemad Area: Anomaly Detection / Geoscience / Self-Supervised Learning Keywords: Geochemical anomaly detection, mineral exploration, benchmark dataset, self-supervised pretraining, Transformer

TL;DR¶

This paper introduces GeoChemAD, the first open-source multi-region multi-element geochemical anomaly detection benchmark (8 subsets covering three sampling media—sediment/rockchip/soil—and four target elements—Au/Cu/Ni/W), and proposes GeoChemFormer, a two-stage Transformer framework that first learns spatial context and then models inter-element dependencies, achieving a mean AUC of 0.7712 that surpasses all baselines.

Background & Motivation¶

Geochemical anomaly detection is a core step in mineral exploration: surface-sample concentrations deviating from the regional baseline may indicate mineralization. Existing research is constrained by two major issues: (1) the vast majority of methods are evaluated on a single-region setting (typically gold-deposit data from the China Geological Survey), making model generalizability impossible to assess; (2) nearly all datasets are closed-source or proprietary, preventing reproducibility and fair comparison. Furthermore, although unsupervised methods offer advantages in generalizability, the anomalies they detect may be unrelated to the target mineralization elements.

Core Problem¶

How to construct a standardized and reproducible benchmark for unsupervised geochemical anomaly detection? How to design an unsupervised detection framework that is both spatially context-aware and capable of distinguishing anomalies relevant to target mineralization elements?

Method¶

Overall Architecture¶

GeoChemFormer operates in two stages: Stage 1 (Spatial Context Learning, SCL) learns latent representations of local geochemical co-variation patterns; Stage 2 (Element Dependency Modeling) detects anomalies conditioned on the learned spatial context.

Key Designs¶

GeoChemAD Dataset: Sourced from publicly available government data released by the Geological Survey of Western Australia (GSWA), comprising 8 subsets—2 sediment (sed1/sed2), 3 rockchip (rock1/2/3), and 3 soil (soil1/2/3)—covering areas ranging from 6 km² to 8,500 km², with sample counts from 224 to 21,040, corresponding to four target mineralization elements: Au/Cu/Ni/W. Each subset provides geochemical concentration CSVs and known mineralization site CSVs.
Stage 1 — Spatial Context Learning: For each query location \(p_i\), its \(K\) nearest neighbors are retrieved via a KD-tree. A token sequence \(\mathcal{S}=[e, q_i, t_1,...,t_K]\) (target-element token + query-location token + neighborhood tokens) is constructed and encoded by an \(L\)-layer Transformer encoder to produce the spatial context representation \(q_i'\). The self-supervised objective is to predict the target-element concentration at the query location from its neighborhood alone, with loss \(\mathcal{L}_{sc} = \frac{1}{N}\sum({\hat{y}_i - y_i})^2\).
Stage 2 — Element Dependency Modeling: The input sequence is \(\mathcal{S}'=[W_g q_i', u_1,...,u_c]\) (geo-context token + \(c\) element tokens). Each element token is formed by concatenating an element identity embedding with the scalar concentration value and projecting the result. A Transformer encoder learns inter-element dependencies conditioned on spatial context. The anomaly score is the reconstruction MSE over all elements: \(s_i = \frac{1}{C}\sum_{c}(x_{i,c}-\hat{x}_{i,c})^2\).
Data Preprocessing Module: Handles missing values (e.g., −9999), the closure problem (CLR/ILR transforms), feature selection (manual/PCA/causal discovery/LLM-assisted), and spatial interpolation (IDW/Kriging).

Loss & Training¶

Stage 1: MSE regression loss (predicting target-element concentration)
Stage 2: MSE reconstruction loss (reconstructing all element concentrations)
Implemented in PyTorch; trained on a single NVIDIA RTX A6000 (48 GB)
Evaluation metric: AUC (averaged over 20 random background samplings)

Key Experimental Results¶

Method Category	Representative Method	Mean AUC
Statistical	Z-score / Mahalanobis / KNN	0.50–0.58
Classical ML	Isolation Forest / One-Class SVM	~0.58–0.62
Deep Generative	AE / VAE / VAE-GAN	0.70–0.73
Transformer	Vanilla Transformer (T1)	0.7147
GeoChemFormer (T2)	Ours	0.7712

GeoChemFormer achieves the best performance on 6 of 8 subsets. On sed2, IF and OSVM perform better (AUC 0.77–0.80 vs. 0.73); on rock2, AE outperforms GeoChemFormer (0.92 vs. 0.83). This indicates that the optimal method varies across geological settings.

Ablation Study¶

SCL pretraining epochs: rock2 peaks at 20 epochs (0.919), sed1 at 40 epochs (0.743), and soil2 requires 60 epochs (0.821).
Neighborhood size \(K\): sed1 improves monotonically as \(K\) increases (16→256, 0.591→0.720), while rock2/soil2 peak at \(K\)=32/16 and then decline—indicating that sediment anomalies benefit from broad spatial context, whereas rockchip and soil anomalies require compact local neighborhoods.
Closure transform: ILR achieves the best mean performance (0.6788), followed by CLR (0.6771), with raw features performing worst (0.6406).
Feature selection: LLM-assisted selection yields the best overall results (0.7412 vs. 0.6419 for manual), while PCA achieves the highest score specifically on GeoChemFormer (0.9427).
Interpolation method: Performance is primarily driven by data characteristics rather than model architecture.

Highlights & Insights¶

GeoChemAD is the first open-source geochemical anomaly detection benchmark spanning multiple regions and sampling media, filling a critical gap in standardized evaluation for this field.
The two-stage Transformer design is elegant: SCL enables the model to capture geographic environmental context, while element dependency modeling realizes target-element-aware anomaly detection.
The paper systematically evaluates the impact of diverse preprocessing strategies (closure transforms, feature selection, interpolation methods), providing practical guidance for real-world applications.
The effectiveness of LLM-assisted feature selection in the geoscience domain is a noteworthy and interesting finding.

Limitations & Future Work¶

The dataset originates from a single geographic source (Western Australia); despite covering multiple sub-regions, geological background bias remains.
The spatial sparsity and irregular sampling inherent to geochemical data constrain the performance of deep learning methods.
On several subsets, simpler methods (IF/AE) outperform GeoChemFormer, indicating that its general advantage is not overwhelming.
Evaluation relies solely on AUC; in-depth analysis of spatial localization accuracy is absent (distance-based metrics are only demonstrated in a case study).

Yu et al. (NRR 2024): Applies Transformers to geochemical anomaly identification but lacks self-supervised pretraining and a standardized benchmark.
Luo & Zuo (Math Geosci 2025): Uses causal discovery algorithms for gold-deposit anomaly detection, but relies on closed-source, single-region data.
Scheidt et al. (NRR 2025): Employs masked autoregressive flow for Li-Cs-Ta exploration with publicly available data, though the method is complex.

Rating¶

Novelty: ⭐⭐⭐⭐ (first standardized open-source benchmark + target-element-aware two-stage framework)
Experimental Thoroughness: ⭐⭐⭐⭐⭐ (systematic comparison across 12 baselines × 8 subsets × multiple preprocessing strategies)
Writing Quality: ⭐⭐⭐⭐ (clear structure, thorough dataset description)
Value: ⭐⭐⭐ (cross-domain application; limited direct value to the CV community, but the dataset contribution is significant)