A Two-Layer Framework for Mineral Prospectivity and Detection Bias in the Canadian Shield
Separating Where Minerals Exist from Where They’ve Been Found
00 — Abstract
This paper presents a Bayesian spatial framework for mineral discovery that explicitly separates geological prospectivity from observational coverage. The two-layer generative model estimates a truth layer (where mineral deposits are likely to exist) and a detection layer (where exploration activity has been sufficient to find them). Validated on orogenic gold deposits across the Ontario Shield.
The framework integrates multi-domain datasets across 1M+ km², automatically selects informative features from 48 candidates, and produces a ranked inventory of 646 frontier cells — areas where geological conditions are favorable but historical examination has been disproportionately low.
Key Findings
Research Outcomes
lift at 5% area screened — the model concentrates known deposits into a fraction of the search space.
of known gold deposits captured within the top 20% of model scores, validating the framework against ground truth.
frontier cells identified — areas of high geological prospectivity that have received disproportionately low historical examination.
Contents
Seven Chapters
The Discovery Problem
Why conventional prospectivity mapping conflates geology with exploration history. The case for separating where minerals exist from where they have been found.
Two Questions, One Model
Architecture of the two-layer generative framework. Separating the truth layer (prospectivity p) from the detection layer (observability q) within a unified Bayesian model.
Data Architecture
Integration of multi-domain datasets — geological maps, geochemical surveys, geophysical grids, and remote sensing imagery — harmonized to a uniform spatial grid at provincial scale.
The Truth Layer
Modelling geological prospectivity. How the system estimates the probability that a cell contains mineralization, independent of whether it has been examined.
The Detection Layer
Modelling observational coverage. Quantifying where exploration has been sufficient to detect deposits and where systematic gaps persist in historical coverage.
Automatic Feature Selection
From 48 candidate geological variables to predictive subsets. Data-driven identification of the ~8 features that carry the most information for each target environment.
Validation & Performance
Model validation against known gold deposits. 19.6× lift at 5% area screened. 98.2% of known deposits captured within the top 20% of model scores.
Frontier Cells
The 646 high-priority targets. Areas where geological prospectivity is high and observational coverage is low — the sweet spots that conventional methods systematically miss.
Framework
Framework Architecture
The system operates across four integrated stages — from raw data integration through two-layer modelling to frontier cell identification.
Fig. 1 — Geological cross-section with targeting overlay
Layers draw progressively as you scroll through each methodology phase.
Data Foundation
Multi-domain datasets spanning geology, geochemistry, geophysics, and remote sensing — harmonized to a uniform spatial grid across 1M+ km² of the Canadian Shield.
Two-Layer Modelling
A generative Bayesian framework that separately estimates the truth layer (geological prospectivity) and the detection layer (observational coverage) within a single unified model.
Feature Discovery
Automatic selection from 48 candidate geological variables to the ~8 most informative for each target system. The model determines what matters, not the operator.
Frontier Ranking
Cells scored by the intersection of high prospectivity and low prior examination. 646 frontier targets identified — structured and ranked for capital deployment.
The system exists regardless of outcome.—
Scale
The Dataset
0+
Square kilometres modelled
0
Integrated datasets
0
Frontier cells identified
0%
Deposit capture rate
01 — The Discovery Problem
Why conventional prospectivity mapping conflates geology with exploration history
Most mineral prospectivity models are trained on known deposits. They learn the geological signatures of ground where mineralization has already been found, then predict more ground that looks the same. This is not prediction — it is recapitulation of detection history.
The fundamental problem: known deposit locations are not a random sample of all deposits that exist. They are a biased sample of deposits that happen to have been found — because someone explored there, because the deposit was shallow enough, because the right technique was applied. Conventional models inherit this bias and amplify it.
Our framework addresses this directly. By separating the question of geological prospectivity (where minerals are likely to exist) from the question of observational coverage (where exploration has been sufficient to find them), we identify ground that is geologically favorable but historically under-examined — the frontier that conventional methods systematically overlook.
Limited Distribution
Request Access
The full paper is available by request to qualified operators, institutional investors, and technical partners with demonstrated interest in systematic mineral exploration.
All inquiries are treated as confidential. Response within 48 hours.