Lanhai builds AI for mineral exploration. We are building Lanhai Targeting, a deep-learning model that learns from Australia's open geoscience data to rank where new copper-gold deposits are most likely, with every target traceable to the data behind it. Baseline v0 is live today.
An exploration team gets a ranked grid of cells and, for every cell, the evidence that put it there. A geologist can audit the ranking before anyone drills. Targets are a starting point for field work, not a claim about what is in the ground.
Before training anything, we built the simplest honest baseline: a 0.1° grid over the Lachlan Orogen window, scored by the copper, gold, silver and molybdenum records within 15 km of each cell, excluding the records inside the cell itself. Deposits weigh three, occurrences one. The score is a density prior, nothing more.
Window 146.5–150.2°E · 31.5–35.0°S · Geological Survey of NSW, WFS 2.0, pulled 1 November 2026 · Method and reproduction steps
It does what a prior does: it re-finds the historical goldfields around West Wyalong and Temora. That is the point. A ranking built on where people already dug cannot find new ground, which is why the model in development learns from magnetics, gravity, radiometrics and imagery instead.
No private data is needed to start. Each layer is public; Baseline v0 uses the first one, the model in development uses all of them. Public inputs mean public checks: anyone can re-run a Lanhai ranking against the same records.
| Source | Layers | Licence | Status |
|---|---|---|---|
| Geological Survey of NSW | Mineral occurrence records (MinView, WFS 2.0), seamless geology, DIGS exploration reports, drill-hole data | CC BY 4.0 | Occurrences in Baseline v0 |
| Geoscience Australia | Airborne magnetics, gravity and radiometric grids, Geophysics Reference Data Collection on NCI | CC BY | In development |
| AuScope NVCL | Hyperspectral drill-core mineralogy, nationwide | Open | Next |
| Satellite imagery | ASTER and Sentinel-2 alteration indices | Open | In development |
Weights-of-evidence porphyry copper-gold model for the Macquarie Arc: evidence layers weighted statistically against known deposits. Same public inputs we use.
Lachlan-wide random-forest models over 300 features at 0.01°. For copper, magnetic grids and intrusive boundaries carried the most signal. Presented at the NSW Government's Exploration in the House.
Deep learning on the raw grids: convolutional and transformer models that read the grids directly, with physics-informed inversion added later. That is the product in development.
Prospectivity at 0.01° across a state is an image problem over hundreds of stacked layers. Training that is GPU work; the rest of the pipeline is designed to move there too. Next: convolutional models over the stacked grids, trained with PyTorch and CUDA on NVIDIA A100 cloud GPUs; the feature stack rebuilt with NVIDIA RAPIDS; then physics-informed inversion of magnetics and gravity with NVIDIA PhysicsNeMo.
Grids, vectors and reports pulled from public portals, tiled to one grid.
Live · occurrencesDerivatives, filters and spatial joins over hundreds of layers with cuDF and cuSpatial.
NextConvolutional and transformer models over stacked grids and imagery, evaluated by holding out known deposits.
NextPhysics-informed inversion of magnetic and gravity data, so the model sees structure, not only texture.
NextLater: an assistant over DIGS exploration reports, so each target also cites what earlier explorers wrote.
LaterConcept illustrations of the two things in development: the single 0.01° stack the model reads, and the depth structure that physics-informed inversion recovers from magnetics and gravity. Not model output, not real data; the real Baseline v0 map is above.
A New South Wales company incorporated in August 2024. Lanhai Targeting is the company's product, built for explorers working the Lachlan Orogen.
Founder. Leads the company and the product.