The Technology Gap in Critical Minerals Exploration

The supply problem is an exploration problem

The shift to clean energy needs large amounts of copper, lithium, cobalt, nickel, manganese, graphite, and rare earth elements. The International Energy Agency expects demand for these minerals to rise four to six times by 2040 in net-zero scenarios. (IEA, The Role of Critical Minerals in Clean Energy Transitions, 2021)

The deposits needed to meet this demand are in the ground. The geology suggests the Earth's crust holds far more undiscovered minerals than the reserves we know about today. The problem is not that minerals are scarce. It is the time and cost it takes to find, map out, and permit new deposits as fast as the transition needs.

Taking a new mine from first discovery to first production takes ten to fifteen years in good locations, and longer where permitting is complex. (World Bank, Minerals for Climate Action, 2021) For new supply to be running in the 2030s, the discoveries need to happen now. That calls for more than just more money. It calls for faster and more accurate ways to explore.

Why exploration has not scaled

Global exploration spending has bounced back from the slump of the mid-2010s, but money alone does not fix the problem. The bottleneck is not funding. It is the ability to read geoscience data well enough to pick out the few drill targets that will actually pay off.

The amount of geoscience data has grown enormously. A modern airborne geophysical survey can produce terabytes of data per project. Regional soil-chemistry databases cover whole continents. Satellites map the surface of the planet down to less than a metre. The data needed to find many hidden deposits already exists. The limit is the power to read it, that is, to combine all of it across an area and find the spots where it points to mineralisation.

Reading all this data by hand does not scale. A geologist on a district project can combine a handful of datasets over a small area. Combining twenty datasets across a whole region is not really possible with traditional methods. Yet that is exactly the scale at which critical minerals exploration must work to find the next wave of deposits.

The specific challenges of battery metals geology

Critical minerals bring their own challenges that have long made them harder to find than ordinary gold or base-metal deposits.

Hard-rock lithium sits in spodumene pegmatites. These bodies follow structures and are often narrow, so a widely spaced survey can miss them. Cobalt is almost always a by-product of nickel, copper, or gold systems, and is rarely the main target. Rare earth elements gather in carbonatite and alkaline intrusions. These have a clear geophysical signature, but it can look like barren rock with a similar signature, so careful sorting is needed. Nickel sulphide deposits form at the edges of mafic-ultramafic intrusions. They can be picked up by geophysics, but it takes careful work on the structures to turn a signal into a target you can drill.

Each of these deposit types leaves a typical signature across several datasets, and a machine learning model can learn that signature. The model does not need to understand the magmatic or hot-fluid processes that formed the deposit. It only needs to learn which mix of features appears with known deposits of that type, and then look for the same mix in unexplored ground.

What AI brings to critical minerals specifically

For critical minerals, the benefits of AI prospectivity mapping are even stronger than for ordinary metals, for three reasons.

First, many regions with critical minerals have plenty of public geoscience data but little exploration history. Government surveys in Canada, Australia, the United States, and parts of Africa have built up wide airborne geophysical and regional soil-chemistry coverage over decades. This data is public and can feed the model even when your own data is thin. The model gains regional context that a geologist on a single project would not usually pull together.

Second, critical minerals often involve deposit types we understand less well than classic gold or copper porphyry systems. The models for some battery-metal deposits are still being refined as new discoveries are made and studied. An AI model that learns straight from the data, rather than from a fixed template, is better placed to catch the full range of signatures for a deposit type we do not yet fully understand.

Third, the money case for AI targeting is clearest where drilling is expensive. In remote northern Canada, Alaska, or Central Africa, a drill hole can cost $150 to $400 per metre, depending on depth, the cost of moving the rig, and the terrain. A well-targeted program that needs fewer holes to hit mineralisation saves money directly, and the saving grows with the size of the program.

The discovery rate problem

Discovery rates have fallen worldwide over the past thirty years, even as exploration spending has gone up. (S&P Global Market Intelligence, 2023) More money now finds fewer and lower-grade deposits. This is partly about geology, since the easiest, near-surface deposits were found first. It is also partly about method. Tools built to find shallow deposits have not kept up with the need to find deeper, fainter, and more complex ones.

AI prospectivity mapping tackles the method side of this problem. By combining more layers of data and spotting fainter mixes of signals than people can catch by hand, it can find deposits that ordinary targeting would miss. These may be deeper, lower in grade, or in settings that do not fit the simple deposit templates in common use.

This does not mean AI will find every hidden deposit. The model is limited by the data on hand and by how varied the target deposits are. But across the huge area of unexplored ground that already has public data, there are almost certainly deposits that today's methods would overlook and a data-driven approach would bring to light.

Where the technology stands today

AI prospectivity mapping is not just a research idea. It is a working tool that exploration companies, government surveys, and junior miners already use to rank drill targets and set budgets. The methods behind it, such as positive-unlabelled learning, spatial cross-validation, and bootstrap uncertainty estimates, are well established. The real question is not whether the technology works. It is whether a given project has the data and the geological context to use it well.

The technology works best together with review by a qualified geoscientist. The model points to statistical patterns, and the geologist judges whether each one means mineralisation or just geological noise. Programs that use both, AI to compare and rank the data and geologists to weigh and choose the targets, get better drill programs than either approach alone.

The gap between the critical minerals the transition needs and the methods we have to find them is real and serious. Closing it means putting the best targeting tools to work on more projects, faster. AI prospectivity mapping is one of the most ready tools we have, and the one with the clearest path from data you already own to better drill decisions.