Machine learning can help mining companies sort decades of geological, geochemical, geophysical and satellite data and decide where to look next. But, argues Iranian scientist and mining engineer Adel Shirazy, an AI-generated prospectivity map is a hypothesis about where to look, not proof of ore, grade, tonnage or economic value. Drawing on a recent copper study in Iran, he explains how to use these tools without mistaking a colorful map for a discovery.
The promise and the trap
Exploration teams often face decades of rock samples, magnetic surveys, satellite images, structural interpretations and records of old workings scattered across multiple sources. Machine learning can organize this evidence and highlight areas resembling known mineralized settings, helping teams compare information consistently and prioritize fieldwork. Reviews of AI in mineral exploration describe the growing use of these methods while emphasizing that their reliability depends on data quality, geological context and model evaluation [1].
The trap is to treat the output as a treasure map. A bright red zone on a prediction map may look precise, but the color is only the result of a model and its assumptions. Unless the score has been properly calibrated, it should not be read as a literal probability that a deposit exists. Nor does a high score tell you the grade, size, metallurgy, depth, permitting difficulty or cost of extraction.
The practical question is therefore not “Can AI find a mine?” It is “Can AI help us decide which targets deserve the next dollar of fieldwork – and how confident should we be in that decision?”
What the model is actually learning
Most exploration models learn statistical relationships between geological features and known mineral occurrences. Depending on the project, inputs may include lithology, faults, alteration, geochemical anomalies, geophysical responses, topography and satellite-derived information. The model uses these patterns to rank areas according to their similarity to known mineralized settings.
Such models can learn genuine mineral-system signals, but they can also learn misleading patterns. If known deposits cluster near roads or historic workings, a model may partly reflect where exploration has already been concentrated. Uneven survey coverage can similarly make data availability appear to be a geological signal.
Comparative research has shown that algorithm choice, training-data distribution, parameter settings and interpretability can all influence mineral-prospectivity results [1], [2]. There is no universally best algorithm for every commodity, deposit type or data environment.
A field-minded approach to AI exploration
A reliable exploration model begins with a geological question, not an algorithm. The commodity, mineral system and scale of investigation should determine which evidence matters. Lithology, structural intersections, alteration and geochemical or geophysical anomalies may all be informative, but their relevance depends on the deposit type. Adding more datasets does not necessarily improve a model; irrelevant layers can introduce noise and obscure meaningful geological relationships.
Data quality must be checked before model selection. Coordinate systems, units, survey dates, spatial resolution, missing coverage, sample preparation, detection limits and the reliability of historical occurrence records can all affect results. No algorithm can compensate for poorly located samples or incompatible datasets. Documenting each layer’s source and limitations helps distinguish geological signals from data-collection artifacts and uneven exploration history.
A model should justify its complexity by outperforming a transparent geological interpretation and a simpler baseline. Evaluation must also reflect real exploration conditions. Randomly splitting nearby samples between training and testing can inflate performance, because neighboring observations may share geological and survey characteristics. Spatially separated test areas, withheld prospects or an independent district provide a more demanding test of whether the model generalizes to unseen ground [3].
From a map to three permits: an example from Iran
A practical example comes from a 2026 study of porphyry copper prospectivity in Iran’s Shahr-e-Babak district, conducted under my supervision. The study combined geological, geochemical and remote-sensing data with an unsupervised autoencoder-based feature-selection framework. Using 14 known mineral occurrences as reference points, the model built on the selected features achieved a prediction rate of 80%, compared with 72% when all available features were used [4]. The resulting map highlighted priority areas for further investigation. After publication, we investigated the identified targets in the field, registered the relevant areas in Iran’s mineral cadastre system and obtained exploitation permits for three high-priority areas. This experience shows how AI can help prioritize exploration targets, while field investigation and the regulatory procedures that follow turn promising predictions into practical outcomes.
A prospectivity map should therefore guide field investigation rather than replace it. High-ranked targets need to be examined through geological mapping, reconnaissance, appropriate sampling and, where justified, geophysical surveys or drilling. Data coverage and uncertainty should be considered alongside predicted prospectivity: a low score in a poorly surveyed area is not strong evidence against mineralization. Field observations, including negative results, should be documented and used to improve subsequent models without compromising independent validation.
Red flags in an AI exploration pitch
Be cautious when a vendor promises “100 percent accuracy”, claims that an AI map proves a deposit exists, or presents results without explaining the data and validation. Before committing funds, ask whether the input layers suit the deposit type, whether historical records have been checked for errors and reporting bias, and whether performance has been tested on spatially independent ground. A credible assessment should disclose data gaps, explain what the scores mean, compare results with a simpler baseline and a geological interpretation, and propose a realistic field test that could confirm or reject the target. If its assumptions, limitations and criteria for revising the recommendation cannot be explained, the map may be more persuasive than informative.
The final test is still geological
A good prospectivity model should make exploration more disciplined by exposing assumptions, comparing evidence and identifying where additional information would be most valuable. It must also be allowed to be wrong. Confidence in a software output is no substitute for independent geological observations, sampling, assays, geophysics or, where justified, drilling.
AI is most useful when it narrows a large search space into a manageable set of testable ideas. The geologist’s job is to decide whether those ideas make geological sense; the field team’s job is to test them. The goal is not a prettier map. It is a better exploration decision.
References
[1] Yang, F., Zuo, R. & Kreuzer, O. P. (2024). Artificial intelligence for mineral exploration: A review and perspectives on future directions from data science. Earth-Science Reviews, 258, 104941. https://doi.org/10.1016/j.earscirev.2024.104941
[2] Rodriguez-Galiano, V. F., Sanchez-Castillo, M., Chica-Olmo, M. & Chica-Rivas, M. (2015). Machine learning predictive models for mineral prospectivity: An evaluation of neural networks, random forest, regression trees and support vector machines. Ore Geology Reviews, 71, 804–818. https://doi.org/10.1016/j.oregeorev.2015.01.001
[3] Roberts, D. R., Bahn, V., Ciuti, S. et al. (2017). Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography, 40(8), 913–929. https://doi.org/10.1111/ecog.02881
[4] Saremi, M., Hoseinzade, Z., Shirazy, A., Shirazi, A. & Beiranvand Pour, A. (2026). Unsupervised autoencoder-based feature ranking and anomaly detection for porphyry copper prospectivity mapping from multi-source geospatial datasets. Minerals, 16(6), 660. https://doi.org/10.3390/min16060660
About the author
Dr. Adel Shirazy, Iranian scientist, is a mining engineer and researcher working on artificial intelligence and machine-learning applications in mineral exploration, geophysical data interpretation and geological modeling. He has postdoctoral research experience at Amirkabir University of Technology in Tehran. ORCID · Google Scholar
This post is also available in: FR