DORA's recommendation scores are a useful starting point that reflect machine learning priorities, but not geological context. The best feature sets combine both: data-driven scores and your understanding of the deposit, data quality, and exploration context.

For a full explanation of how scores are calculated, see [DORA 2.0] How DORA 2.0 Scores Input Features.

More features are not better. DORA defaults to 64 recommended features, but you do not need to use all available slots. In practice, 20–30 well-chosen layers consistently produce strong models. Often, starting with the top 24 or 32 recommended features is a reasonable approach for an initial experiment.

Don't include the target element as an input feature. Using the same element as both your learning point threshold and an input raster creates circular logic and inflates results without adding genuine predictive value (see Stage 1 below).

A score of zero doesn't mean a layer isn't useful. A common reason for a zero score is partial AOI coverage. If a dataset covers your learning points and property boundary, it may still be worth including.

High scores tend to favour broad, lower-resolution data. Rasters with complete AOI coverage and longer-wavelength signals tend to score well. This makes sense statistically, but can mean that high-resolution local surveys are undervalued. Keep this in mind when reviewing recommendations.

Working through three sequential experiments (adding a new data category at each iteration) can help you identify the most robust targets and understand what each dataset contributes to the prediction.

Targets that appear consistently across all three iterations are your highest-confidence candidates.

In the first stage, use objective, continuous-coverage data only, including geophysical layers and lineaments derived from geophysical products.

Accept DORA's recommendations as your starting point, but exclude any rock rasters that match the target elements used in your learning points, as these create circular logic. Additionally, exclude low QC soil datasets at this stage, as anomalies in these layers may reflect auxiliary grids bleeding through the geochemical prediction rather than genuine signal.

Incorporate geochemical layers and compare results against your baseline.

In stage three, add interpretive geology layers, such as mapped structures, lithological distance maps, and similar features.

Note that not all geology layers carry equal confidence; a precisely mapped fault is a very different input from a regional public geology boundary. Be sure to consider how each layer was produced and whether it is appropriate for the scale of your AOI before including it.

Reach out to your dedicated DORA contact or email support@VRIFY.com for more information.