What Is Mineral Prospectivity Mapping?

The core problem in mineral exploration

Finding a mineral deposit is a spatial prediction problem. A geologist or an algorithm must answer the same question: given everything we know about the geology of this area, where is mineralisation most likely to occur?

The challenge is that the answer is never obvious. Mineral deposits are geologically rare — they form when a specific combination of heat sources, fluid pathways, chemical gradients, and structural traps converge in exactly the right way over millions of years. That combination leaves signatures across many data types: geophysical anomalies, geochemical halos, structural patterns, alteration zones. Interpreting those signatures manually, layer by layer, is slow, expert-dependent, and limited by how many variables a human mind can hold simultaneously.

Prospectivity mapping is the field's systematic answer to this problem. Instead of looking at one dataset at a time, it combines all available evidence into a single surface that expresses the relative likelihood of mineralisation at every point across the area of interest.

How traditional prospectivity mapping works

In the traditional approach, a geoscientist defines a conceptual deposit model for the commodity being targeted. The model specifies which geological criteria are necessary or favorable for that deposit type: proximity to intrusions, structural setting, host lithology, alteration mineralogy, geophysical response. The geologist then scores each criterion across the map and combines the scores — manually or through weighted overlay in GIS — to produce a prospectivity surface.

This approach works, and it encodes decades of geological knowledge. Its limitations are: it is only as good as the deposit model the geologist defines, it does not scale easily to large areas or many input layers, and it treats criteria as independent when in practice they interact in complex ways. A subtle combination of three moderate anomalies across three datasets may be more significant than any single strong anomaly — and that kind of interaction is difficult to capture with manual overlay methods.

What AI prospectivity mapping does differently

AI prospectivity mapping inverts the traditional workflow. Instead of a geologist defining which criteria matter, the model learns the criteria from the data — specifically, from the feature combinations that occur near known mineral deposits or mineralised drill intersections within the dataset.

The model is trained on the locations of known occurrences within the area of interest. It learns which spatial patterns — across geophysics, geochemistry, structural geology, lithology, and remote sensing data simultaneously — co-occur with mineralisation. It then applies that learned pattern to every unsampled point on the map, producing a probability score at each location.

The output is not a yes/no classification. It is a continuous surface where higher values indicate greater statistical similarity to the feature combinations observed near known deposits. A geologist reviewing the surface can identify clusters of high-probability pixels, examine which input layers drove each cluster, and decide whether the underlying geology supports drilling.

The class imbalance problem

The most technically significant challenge in AI prospectivity mapping is class imbalance. Known deposits are rare — often one to three occurrences per district — while non-mineralised terrain makes up the overwhelming majority of the area. A naive classifier trained on this data would learn to predict “no deposit” everywhere, achieving high accuracy while being completely useless.

The field has developed several strategies for this. Positive-unlabelled (PU) learning treats the non-deposit locations as unlabelled rather than negative — acknowledging that some of them may be undiscovered deposits. Ensemble methods train multiple models on bootstrap samples of the data and aggregate predictions, making the result more robust to the sparse positive set. Spatial cross-validation ensures that validation is done on genuinely held-out geographic areas rather than random splits, which would overstate model performance due to spatial autocorrelation.

These are not optional refinements. Without them, a prospectivity model trained on geoscientific data will look good on paper and fail in practice.

What a prospectivity map looks like in practice

A finished prospectivity surface is a raster file — typically a GeoTIFF — where each pixel carries a score between 0 and 1 representing the estimated probability of mineralisation. The surface is continuous: there are no hard boundaries between high and low zones. High-probability pixels tend to cluster, forming target corridors that a geologist can compare against known structural trends and geological interpretations.

Accompanying the surface is an uncertainty map that quantifies model confidence at each location. Uncertainty is high where the training data is sparse or where different model runs disagree. The two maps are used jointly: a location that is both high-probability and low-uncertainty is a defensible drill target. A high-probability, high-uncertainty location is a signal that more data collection — another geophysical traverse, additional geochemical sampling — would be the higher-value next step before committing to drilling.

What prospectivity mapping cannot do

AI prospectivity mapping identifies statistical anomalies — locations where the feature combination is similar to what has been observed near known deposits. It does not understand geological processes. It cannot reason about structural geology the way a senior structural geologist can. It does not know that a geophysical anomaly is explained by a mapped intrusion rather than a sulfide body. In geological settings with no analogues in the training data, performance degrades.

The output is a prioritisation tool, not a drilling decision. Every prospectivity map should be reviewed by a qualified geoscientist before informing a drill program. The model surfaces what the data says; the geologist decides what it means. Used well, AI prospectivity mapping extends the geologist's bandwidth — covering more ground, cross-referencing more data, finding patterns that are invisible at human scale — while leaving judgment in the hands of the expert who understands the geology.