A Blog by Jonathan Low

 

Dec 13, 2022

How AI Is Helping Identify New Sources For EV Battery Materials

More efficient, longer lasting batteries are the key to renewable energy alternatives and rare metals are the key to making those batteries. 

AI is being used not only to identify new sources of those metals, but also to potentially find metallic alternatives not previously considered. JL

Vince Beiser reports in Wired:

As the world begins to shift from fossil fuels to greener alternatives, there’s a global scramble to find the vast quantities of cobalt, lithium and other metals required to build electric car batteries, solar panels, and wind turbines. But finding new mineral deposits has been difficult and expensive, and getting more so. Companies are trying to make the process faster, cheaper, and more efficient by applying artificial intelligence. A database incorporating all the information it about the Earth’s crust—geologic reports, soil samples, satellite imagery, academic research, and handwritten field reports - makes this machine-readable. “Machine learning can pick up patterns in the distribution of elements." “THESE THINGS ARE hard to tip over,” geologist Wilson Bonner assures me as the four-wheeled all-terrain vehicle he’s piloting tilts suddenly sideways, pitching me toward the churned up mud beneath our wheels. We’re grinding up the side of a thickly forested hill in rural Ontario, Canada, on a chilly fall day, heading toward a spot that Bonner’s employer, startup KoBold Metals, says represents the marriage of cutting-edge artificial intelligence with one of humanity’s oldest industries.

 

We do indeed complete the half-hour trek relatively unmuddied, finally breaking through a ring of broken trees and mangled brush into a swath of bulldozed mud. A black pipe about as wide around as my arm juts out of the ground—the top end of a hole nearly a kilometer deep that was punched into the ground by a truck-sized drilling rig that sits idly nearby. It’s not much to look at, but this hole might mark a step into the future of mining, an industry crucial for the world’s transition to renewable energy.        

 

 

 

 

 

 

“KoBold is doing the riskiest thing,” says Sam Cantor, head of product at Minerva Intelligence, another AI-driven mining exploration startup. Even with help from AI, placing bets on potential mineral deposits is far from a foolproof process; metals often turn up in places with wildly different conditions and geologic histories. “When you’re training an algorithm to recognize a face, you can assume there’s a mouth and it’s below the nose and eyes,” Cantor says. “But if you apply that training to insect faces, you might find more than two eyes and no nose. Training an algorithm on data from Alaska and applying it to Nevada means it might have a lot of wrong assumptions.” But the payoff from a big find can be stupendous. Earlier this year Tesla agreed to buy $1.5 billion worth of nickel from a new mine in Minnesota slated to open around 2026.

Copper and nickel were previously found at the Crystal Lake site KoBold is now exploring back in the 1970s, but not in high enough concentrations to make mining profitable. The startup’s algorithms, however, suggested there may be more there. So the company sent in a team of geologists and technicians, headed by Bonner, to gather more data. They circled the targeted hill with a couple of miles of yellow electric cable, ran a current through it and logged where the current generated a magnetic field underground. This electromagnetic survey found seven or eight potential deposits, but the team didn’t know for sure whether they were copper or nickel, or something else altogether, like graphite. Nor did they know the exact shape, size, or location of those deposits. A small one close to the surface, for instance, can have the same electromagnetic signature as a large one deeper down.

Once again, KoBold turned to algorithms. Finding out exactly what’s underground requires drilling, but that is time-consuming and expensive, and it requires tearing up land, all of which KoBold would like to keep to a minimum. So from her home in Boulder, Colorado, KoBold data scientist Beth Reid deployed a machine-learning system, based on a more general version first developed at Stanford University, to generate models of the thousands of different configurations of underground minerals that could have caused the electromagnetic readings picked up in Ontario. Bonner used his geology experience and intuition to help screen out unlikely suggestions. Reid then worked to figure out how to drill a single hole that would narrow down those possibilities as much as possible—that is, what precise location, depth, and angle would intersect the largest number of all the possible deposits, proving or disproving which ones are actually there. On the ground at the Crystal Lake site, Bonner then applied those calculations to position the drill. The result: that hole in the muddy clearing.

ADVERTISEMENT

In theory, that single hole will provide more information than a dozen poked into the ground with traditional methods. At the time of my visit, though, the team still didn’t know if it had turned up anything. They have to wait until the cylindrical rock samples they’d brought up come back from the lab where they were sent for chemical analysis. Even if they didn’t strike metal, though, the samples will at least provide another layer of data offering a fresh set of clues. “Machine learning can pick up patterns in the distribution of elements, which informs our understanding of what’s down there,” says Reid. “It all helps determine where next to drill.” 

While Kobold’s technology may make the exploration process more efficient, it still doesn’t guarantee that anything will be found. “It’s the explorer’s dream, to be told exactly where to drill, but we haven’t seen that yet from any of these systems,” says Mathieu Landry, a Canadian geoscientist who consults with mining companies. He recently coauthored an article in the journal of the Society of Economic Geologists that concluded that the impact of AI “on actual business success—in this case measured in terms of ore deposit discovery—is far from certain.” The article added: “AI has a long history of overpromising and under-delivering.”

Landry thinks AI is more likely to be useful to miners for narrower tasks like analyzing elements in rock samples than for searching the whole planet. In any case, even if KoBold does find copper and nickel in Crystal Lake, it will take several years before any of it hits the market. More certain is that if AI can speed up any part of the process of finding new mineral deposits, it will be a welcome boost in the race for the crucial metals needed to decarbonize our lives.

0 comments:

Post a Comment