GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration
| Authors | Yihao Ding et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2603.13068 |
| Download | |
| Categories | cs.LG, cs.AI |
Abstract
Geochemical anomaly detection plays a critical role in mineral exploration as deviations from regional geochemical baselines may indicate mineralization. Existing studies suffer from two key limitations: (1) single region scenarios which limit model generalizability; (2) proprietary datasets, which makes result reproduction unattainable. In this work, we introduce \textbf{GeoChemAD}, an open-source benchmark dataset compiled from government-led geological surveys, covering multiple regions, sampling sources, and target elements. The dataset comprises eight subsets representing diverse spatial scales and sampling conditions. To establish strong baselines, we reproduce and benchmark a range of unsupervised anomaly detection methods, including statistical models, generative and transformer-based approaches. Furthermore, we propose \textbf{GeoChemFormer}, a transformer-based framework that leverages self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples. Extensive experiments demonstrate that GeoChemFormer consistently achieves superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability. The proposed dataset and framework provide a foundation for reproducible research and future development in this direction.
Engineering Breakdown
Plain English
GeoChemAD is an open-source benchmark dataset for geochemical anomaly detection in mineral exploration, addressing a critical gap where existing models fail to generalize across regions because they're trained on proprietary, single-region datasets. The authors compiled eight diverse subsets from government geological surveys covering multiple spatial scales and sampling conditions, making results reproducible. They then benchmarked a range of unsupervised anomaly detection methods—statistical models, generative models, and transformer-based approaches—to establish strong baselines for the community. This work solves the reproducibility crisis in mineral exploration ML by providing both the standardized data infrastructure and comparative baselines that the field lacked.
Core Technical Contribution
The primary contribution is GeoChemAD itself: a multi-region, multi-source geochemical dataset that moves the field beyond single-region silos and proprietary black boxes. Beyond the dataset, the authors establish a rigorous benchmarking framework that compares unsupervised anomaly detection methods—including statistical baselines, generative models (likely VAEs or normalizing flows), and transformer-based architectures—on a standardized evaluation protocol across diverse geological contexts. This is novel because prior work either used proprietary data, limiting reproducibility, or tested on single regions, preventing assessment of cross-region generalization. The systematic comparison across method families (statistical, generative, attention-based) on a realistic, government-sourced geology task is the missing piece that enables the community to make informed method choices.
How It Works
The system takes geochemical measurements from multiple regions (element concentrations, spatial coordinates, and sampling metadata) and learns anomaly detection models without labeled mineralization events. For each region, the input is a multivariate geochemical vector representing elemental compositions; the model learns what 'normal' baseline geochemistry looks like for that region by analyzing the statistical distribution or learned latent manifold of the data. The benchmarked approaches include: (1) statistical methods (likely Isolation Forest, Local Outlier Factor, or Mahalanobis distance) that flag samples deviating from regional statistical norms; (2) generative models (VAEs or autoencoder-based) that reconstruct normal samples and flag high reconstruction error as anomalies; (3) transformer-based methods that learn contextual attention patterns over elements and spatial features to identify deviations. The output is an anomaly score for each sample, enabling mineral exploration teams to prioritize drilling locations or further surveying based on geochemical deviation magnitude.
Production Impact
For mineral exploration teams, this enables cost-effective screening of geochemical survey data without expensive manual expert review or region-specific model retraining. An exploration company can now download GeoChemAD, benchmark their preferred anomaly detection method on it (taking minutes to hours on commodity hardware), then apply that method to their own regional data with confidence that the approach generalizes, rather than guessing between untested proprietary black boxes. The trade-off is that unsupervised methods will produce some false positives (flagging barren anomalies that aren't economically valuable minerals), requiring geologists to validate—but this reduces false negatives (missing real mineralization) compared to purely statistical rule-based systems. Integration is straightforward: wrap your geochemical assay pipeline to produce element-concentration vectors, feed them through a deployed anomaly detection model, and surface high-anomaly regions to drilling teams. The benchmark comparison lets teams choose methods balancing speed (statistical models run in milliseconds on edge devices) versus sensitivity (transformer models catch subtle multi-element patterns but require GPU at inference).
Limitations and When Not to Use This
The paper assumes geochemical anomalies cleanly correlate with economic mineralization, which is not always true—many geochemical deviations reflect barren geological noise, so false-positive rates will be high without domain expertise to filter results. It does not address temporal dynamics (geochemical weathering, seasonal variation in sampling methods) or spatial autocorrelation (samples collected near each other are not independent), which could bias baseline estimates if the data collection process changed over time or across sub-regions. The 'multi-region' benchmark may still be geographically and lithologically biased toward the geology of the surveyed areas; models trained on granitic terrains may fail on sedimentary or volcanic settings not well-represented in the eight subsets. Finally, the paper is incomplete in the abstract—it mentions reproducing methods but does not detail results, making it impossible to assess which method families actually work best or what accuracy is achievable, limiting immediate practical guidance for practitioners choosing between approaches.
Research Context
This work sits at the intersection of applied ML and Earth sciences, building on decades of statistical geochemistry (regional baseline/threshold analysis) while bringing modern unsupervised learning techniques (VAEs, attention mechanisms) to a domain that has historically relied on manual expert interpretation. It mirrors similar open-source benchmark efforts in medical imaging (ImageNet for histopathology) and climate science (large-scale multi-site climate datasets), recognizing that reproducible benchmarks are foundational for method development. The paper directly responds to critiques in the anomaly detection literature that most benchmarks test on tabular data from finance or manufacturing, not the geochemical timeseries and spatial data relevant to Earth science; GeoChemAD fills this gap. This likely opens a research direction in geochemical anomaly detection where transformer-based methods and meta-learning (adapting models across regions) become viable, similar to how multi-dataset benchmarks accelerated progress in medical image analysis.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::