AI Perception Gap: How AI Is Revolutionising The Hunt For Critical Battery Metals

The Problem

The race towards a greener future — dependent on battery production and the electric vehicle (EV) revolution — hinges on four critical elements from the periodic table: Copper, Lithium, Cobalt, and Nickel. But why these four?

• Copper — the workhorse of a battery — powers the flow of electricity by carrying electrons in almost all electricity applications. An EV requires three times as much copper (by mass) compared to an equivalent internal combustion car (a petrol-fuelled vehicle).
• Lithium is the most desired anode material in a battery. It’s the lightest metal and has a high propensity to lose an electron, which is necessary to produce energy within the battery. Alternatives like sodium-ion batteries are far less effective in comparison: sodium’s mass is three times that of lithium, and its reluctance to lose electrons results in significantly less energy per unit of mass. Simply put, lithium is the lifeblood of EV batteries!
• Cobalt and nickel are the best cathode materials since their chemical properties make them great receivers of electrons. While cobalt delivers peak battery performance, nickel has emerged as the preferred substitute due to cobalt mining practices. Over 60% of the world’s cobalt supply comes from mines in the Democratic Republic of Congo, where many mines force children in their early teens to work in horrendous conditions. Furthermore, Russia is the second biggest producer of both cobalt and nickel. The ethical and political urgency of discovering alternative deposits of both metals is undeniable.

But here’s the crux: we don’t have enough metal to meet skyrocketing demand. Experts predict a $15 trillion supply gap in these four metals by 2050 (based on March 2023 commodity prices). Yes, that’s $15 trillion worth of new discoveries needed (after we’ve mined all known deposits first).

By 2050, demand is forecasted to explode — 30x for lithium, 15x for cobalt, 5x for nickel, and 2x for copper (a massive increase considering copper is already a $100bn market).

Meanwhile, supply is dwindling. Imagine fishing in the same pond for decades. Eventually, the big fish run out. The small ones aren’t economically worthwhile. You’re forced to either stop fishing or find a new pond.

That’s where we are with metals. For centuries, we’ve mined just the top one-hundred meters beneath the surface because these deposits are easier to discover and extract. Now that these deposits have diminished, we’re left with no choice but to dig deeper. But, our old discovery techniques no longer suffice in this uncharted territory; we need new methods to fish for metals at greater depth.

Enter KoBold Metals, a Silicon Valley-based startup founded in 2018 to tackle the bottleneck threatening the battery metals revolution. Their mission? To revolutionize mineral exploration by looking hundreds, even thousands of meters below Earth’s surface. And the key to their strategy? AI-powered exploration.

Exploration in a Toy Example

Before we focus on all things KoBold, let’s begin with a simple example…

Aim

Your goal is to create a map of a region no one has ever visited before. You must predict the class (sea, forest, sand dune, etc.) of the terrain at various points. All you have at your disposal is a plane, 4 parachute jumpers, and devices enabling the crew to communicate.

Prior Knowledge

You begin by flying the plane at 10,000m; traversing the whole region like a counter advancing through a snakes and ladders board. At this altitude, your ability to make confident predictions is limited. Hence, you assign subjective probabilities to regions: “I predict the big blue region is the sea and I am 98% confident.” Here, you made a prediction (sea) and indicated your confidence in your prediction (98%). Are the white regions snow, sand dunes, or salt flats? White regions immediately beside the sea are more likely to be sand dunes than anything else. Moreover, snow is more likely to be found at higher altitudes, so you look for mountains when distinguishing from salt flats. After flying over the entire region, you create a first draft of the map with your team. In some parts you’re very confident, others you’re not. How do you increase the confidence in your observations?

Field Mapping

Of course, you drop the 4 parachuters! As they fall, the radius of their visual field shrinks but they become more confident in their predictions. We call an increase in confidence a reduction in uncertainty.

Why keep each parachuter’s belief separate from one another? If all 4 parachuters can all see the same green region during their fall and they all predict it is a grassland, your team can confidently predict grassland on the map. If you spread the parachuters over the entire map and each region can only be observed by 1 team member, there would be more uncertainty in each of your predictions. Now, a trade-off exists: spread the parachuters out (resulting in non-overlapping views) to cover a more expansive area but have less confidence in your conclusions, or cover less area but have more confidence in your conclusions. Does an optimal strategy exist?

Most likely! Prior to dropping the parachuters, you can use the draft map of the world (called your prior knowledge). Recall, your prior is confident that the blue region visible from the plane is the sea. Is it wise to drop parachuters there? No, of course not! You’re already confident after all. Dropping your team where uncertainty is greatest (confidence is low) is optimal. You want each resource (each parachuter) to provide as much new information as possible.

The overarching principle is: build a prior knowledge of the world using all tools at your disposal, then decrease your uncertainty as much as possible through exploration with the limited resources you have.

Exploration at KoBold Metals

Aim

KoBold Metals’ aim is to efficiently use their resources to create a highly informative map. In their case, their map must indicate the presence of metals beneath the surface.

Prior Knowledge

In the previous example, the map represented visible light and the only measuring device employed was the human eye. Of course, KoBold use tools far more capable than the human eye and they use data types far more informative than visual light. The types of data they collect to build their prior knowledge are:

• Spectral imagery: Spectral imagery can be visible light, UV light, infrared light, and more.
• Magnetic: Anomalies in magnetic maps that cannot be explained by Earth’s magnetic field alone can indicate the presence of metal.
• Radiometrics: Rocks are continuously decaying and emitting radioactive decay products. Certain products can indicate the presence of certain metals.
• Electromagnetic: Provides information about the conductivity of the surface. An electromagnetic sensor emits electromagnetic waves into the ground or rocks. These waves induce currents in conductive materials beneath the Earth’s surface and the sensor measures the secondary electromagnetic fields generated by these currents.
• Gravity: Anomalies in gravity maps that cannot be explained by Earth’s gravitational field alone can indicate the presence of metal. Density of materials beneath the surface influence gravity readings above the surface because denser material creates a stronger gravitational attraction.

This stage of the process is absolutely critical. When using AI, the real power lies in the data itself. AI may be a brilliant detective, capable of uncovering patterns present in large quantities of data, but it can only work with the clues it’s given. KoBold’s job? Plant those clues. By ensuring the data is rich, precise, and comprehensive, they lay the groundwork for AI to transform raw information into groundbreaking discoveries.

Field Mapping

KoBold is searching for metals across five continents, but here I focus on their exploration at a site in Quebec, Canada. The Cape Smith Belt, Quebec, is a complex geological structure comprised of a variety of rocks over 1.9 billion years old. It is an extremely remote and beautiful area, making it a popular destination for hiking, camping, and fishing. The remote location imposes immense challenges in terms of logistics and cost. Everything — drilling equipment, personnel, survival supplies, etc. — must be flown in. Moreover, all samples taken by drilling into the rock must be flown out for testing. The climate presents further challenges: there’s snow on the ground nine months a year, leaving only three months to collect samples.

Across a 1000 km² (200,000 acres) region searched by KoBold Metals, only 5 tonnes of samples were collected during an intensive three-month campaign. To put this in perspective: the rock samples gathered amount to the mass of just four hatchback cars, collected from a region comparable to 98,000 football (soocer) pitches — crazy! Hence, it is vital to be highly efficient. Explorers must collect the most informative type of information in the most informative locations.

In the image below, we see the prediction and uncertainty of KoBold’s prior knowledge on the left (Day 1 prediction and Day 1 uncertainty). After drilling at four locations (shown by four black dots in Day 1 uncertainty), KoBold updated their prediction and uncertainty, shown in Day 30 prediction and Day 30 uncertainty. They classify each region into one of three classes (Basalt, Gabbro, Peridotite), similar to spotting sand dunes, salt flats, and sea in the toy example. Here, each class is represented by green, blue, or red, as shown by the prediction plots.

As in the toy example, where you chose not to drop parachute jumpers in an area you were highly confident was the sea, KoBold Metals were guided by the uncertainty in their prior knowledge. The drilling locations they selected, shown by the four black dots in Day 1 uncertainty, clearly lay in regions of high uncertainty (areas of yellow). Their confidence at these very locations improves tremendously, as evidenced by the low-uncertainy blue on the Day 30 uncertainty plot. Furthermore, the dominant colour of the uncertainty plot switches from yellow to green after the field mapping, suggesting uncertainty has been reduced throughout the region. This decrease in uncertainty is also evident on the prediction plots. The Day 30 prediction plot appears to be a less blurry version of the Day 1 prediction plot, with well-defined areas of blue, red, and green. Essentially, a reduction in uncertainty provides, quite literally, a clearer image of the minerals in the exploration region.

The Big Picture

I have simplified the entire exploration process significantly here. There are many scientific challenges KoBold face throughout their search for mineral deposits. To learn more about the end-to-end process, watch the video below or listen to Kurt’s appearance on a podcast here.

Have KoBold Metals Been Successful?

Yes! In 2024, they announced their first major discovery: a colossal copper deposit in Zambia. This article states it’s the largest copper deposit ever recorded in Zambia and this article suggests the mine contains enough copper to produce 100 million of today’s average-size electric vehicle batteries.

I think stories like this are so incredible. Kurt House launched a startup to approach mineral exploration differently. They built their workforce to be comprised of tens and tens of data scientists — unprecedented in the industry. Then after years of trusting their process, with collaboration across disciplines between geoscientists and data scientists, they struck gold. I voiced my admiration for their discovery when I met Kurt during my trip to Stanford in June 2024. His response was typical of a leader on a mission: “Thanks. Onto the next.”

A Note on Sustainability

I know what the sustainability-minded amongst us are thinking…support the green revolution by building additional mines and further disrupting the beautiful planet. Some contradiction. Here’s the situation…

Drilling for fossil fuels and burning them to produce energy is a singular economy. Carbon is released into the atmosphere where it remains for a significantly long time. Whereas, drilling for metals to produce EV batteries can facilitate a circular economy on a gigantic scale. It’s possible to achieve a 98% or more recycling rate on the metals contained in batteries; high enough to almost stop mining metal altogether! But, we must extract the metals from the ground and funnel them into the system first. This will take 50 years or so — hold tight!

While the extraction of battery metals is a positive step for green energy and combating climate change, the term sustainability encompasses more than just carbon emissions. Mining disrupts ecosystems, which no amount of reduction in emissions can bring back. Thus, KoBold’s work doesn’t have a wholly positive impact. This is where my friends with academic backgrounds in philosophy and environmental studies can take over…bye for now.