Traditional exploration means expensive drilling campaigns with uncertain outcomes. This AI-driven approach reads the data a company already has (seismic surveys, geological maps, geochemistry) and ranks prospects before a single drill bit turns.
The challenge
Sparse, heterogeneous geological datasets spanning decades of collection methods. Modelling complex 3D subsurface structures from 2D survey data. Quantifying prediction uncertainty well enough to guide a capital-heavy drilling decision. And validating against known reservoirs while still generalising to new basins.
The approach
A 3D convolutional architecture built for volumetric seismic interpretation, with geological priors enforced through physics-informed loss functions and thermodynamic constraints. A Bayesian uncertainty layer attaches a confidence interval to every prediction, so each prospect ships with a calibrated "how sure are we."
Predictions that obey the physics of the subsurface are what let a team bet a drilling budget on the model.
The results
Three previously unknown geothermal prospects were identified in test regions, and two were confirmed by subsequent drilling.
- 85%
- Detection accuracy
- 60%
- Lower exploration cost
- 18 → 3 mo
- Survey to prospect
Exploration costs fell 60% versus traditional methods, and survey-to-prospect time dropped from 18 months to three. The technology now runs across multiple geothermal exploration programs.