New Machine Learning Framework Fuses Hyperspectral and Geochemical Data for 3D Geological Modeling
At the European Geosciences Union (EGU) General Assembly 2026, researchers from Finland’s Geological Survey (GTK) and Aalto University presented a novel machine learning framework that fuses two complementary data types: drill‑core hyperspectral imaging and traditional geochemical assays. The goal is to produce high‑resolution 3D geological models that reveal not just rock types but also alteration halos and mineralization vectors — critical information for gold exploration.
The framework was tested on a well‑known orogenic gold system in central Finland, an area with decades of drilling data but still underexplored at depth. Hyperspectral scans captured mineralogical variations (e.g., white mica, chlorite, carbonates) every few centimeters along drill cores. Geochemical data provided element concentrations (Au, As, Sb, etc.) at variable intervals. Conventional modeling kept these layers separate. The new ML framework — based on a self‑supervised transformer architecture — aligned and integrated both datasets into a single 3D voxel model.
Results were striking. The integrated model correctly identified the main ore shoot and predicted two blind zones that earlier models missed. In one case, the framework detected a subtle shift in white mica chemistry 200 meters above a gold intercept — a vector that experienced geologists would call “the halo effect.” Because the ML model learned this pattern from data, it could quantify uncertainty and suggest optimal locations for new drill holes.
“We are moving from ‘geological intuition’ to ‘geological intelligence’”, said Dr. Henna Lahti, lead author. “The framework doesn't replace geologists — it augments their ability to see patterns across scales.”
The team has released the code under an open license, along with a synthetic training dataset. Several mining companies have already expressed interest in applying the framework to their own projects, particularly in brownfield exploration where hundreds of existing drill holes can be re‑analyzed without additional drilling. This aligns perfectly with the industry’s push toward more sustainable, lower‑impact exploration — extracting more knowledge from each meter drilled rather than drilling more meters.
The next milestone: real‑time integration with automatic drilling rigs, creating a closed loop where data collected at 10 AM refines the 3D model and guides drilling decisions by 2 PM the same day.
The framework was tested on a well‑known orogenic gold system in central Finland, an area with decades of drilling data but still underexplored at depth. Hyperspectral scans captured mineralogical variations (e.g., white mica, chlorite, carbonates) every few centimeters along drill cores. Geochemical data provided element concentrations (Au, As, Sb, etc.) at variable intervals. Conventional modeling kept these layers separate. The new ML framework — based on a self‑supervised transformer architecture — aligned and integrated both datasets into a single 3D voxel model.
Results were striking. The integrated model correctly identified the main ore shoot and predicted two blind zones that earlier models missed. In one case, the framework detected a subtle shift in white mica chemistry 200 meters above a gold intercept — a vector that experienced geologists would call “the halo effect.” Because the ML model learned this pattern from data, it could quantify uncertainty and suggest optimal locations for new drill holes.
“We are moving from ‘geological intuition’ to ‘geological intelligence’”, said Dr. Henna Lahti, lead author. “The framework doesn't replace geologists — it augments their ability to see patterns across scales.”
The team has released the code under an open license, along with a synthetic training dataset. Several mining companies have already expressed interest in applying the framework to their own projects, particularly in brownfield exploration where hundreds of existing drill holes can be re‑analyzed without additional drilling. This aligns perfectly with the industry’s push toward more sustainable, lower‑impact exploration — extracting more knowledge from each meter drilled rather than drilling more meters.
The next milestone: real‑time integration with automatic drilling rigs, creating a closed loop where data collected at 10 AM refines the 3D model and guides drilling decisions by 2 PM the same day.
