Machine learning discovers important metal deposits for making electric vehicle batteries.

In June 2022, six Boeing 737 aircraft loaded with tents, food, satellite internet equipment, drones, geophysical survey equipment, drilling equipment, and a team of experienced geologists flew to a remote, rudimentary tarmac in northern Quebec, Canada. Geologists are looking for major mineral deposits needed for clean energy in the future. Cutting-edge scientific computing is a blend of old-school risk-taking, and they seem to have the help of both Alan Turing and Indiana Jones. Our start-up KoBold Metals owns 800 square kilometres of mineral rights in this part of Canada, and this determination is based in part on the predictions of artificial intelligence systems. According to artificial intelligence, there is good reason to believe that we will discover valuable nickel and cobalt deposits underground. In this near-Arctic region, summer snowmelt creates a brief window into which a small amount of equipment and personnel can be brought in to validate our predictions.

In 2018, with the support of Bill Gates' Breakthrough Energy Ventures (BEV) and Silicon Valley Venture Andreessen Horowitz, we co-founded KoBold with the goal of finding ways to discover new deposits of important metals needed for electric vehicle batteries, which are in huge and growing demand.

We try to transform mineral exploration from a process of manual exploration, judgment-guided, and trial-and-error to a data-driven and scalable science. At depths of 100 to 2,000 meters, it is more difficult to find valuable deposits of cobalt, copper, lithium and nickel than to find a needle in a haystack.

Preventing the catastrophic impacts of climate change requires net-zero greenhouse gas emissions by 2050, including replacing all fossil fuel-powered light cars and trucks with electric vehicles. As a result, billions of EV batteries will need to be manufactured. Even today, demand for metals outstrips supply, as evidenced by the doubling of nickel prices in 2022 and the five-fold increase in lithium prices. To enable the global transition to electric vehicles, we need to discover and mine $15 trillion worth of cobalt, copper, lithium, and nickel by mid-century. (Our current goal is to mine about $3.6 trillion worth of these metals by 2050).

Leaders around the world and regions are well aware of this need. For example, in March 2022, United States President Joe Biden launched the Cold War-era Defense Production Act, using the powers granted to the president by the law to encourage domestic production of minerals needed for electric vehicle batteries. In August 2022, the Inflation Reduction Act was signed into law, which includes billions of dollars to subsidize the development and operation of metal minerals in the United States and globally.

Investors are also aware of supply-side challenges. In February 2022, KoBold raised $192.5 million in a Series B funding round to acquire more than 50 exploration sites in Australia, Canada, Greenland, sub-Saharan South Africa and United States. We plan to use artificial intelligence to streamline much of the fragmentation process in exploration for new deposits. Once the deposit is discovered, we plan to work with the mining company on the actual mining operation, while again using AI tools to advise them on efficient mining.

As early as thousands of years ago, humans noticed that rocks containing useful minerals had an eye-catching appearance. For example, iron sulfide, the main mineral in nickel sulphide deposits, is exposed to air and rain, which can produce significant red rust. Weathering has made copper sulfide a variety of brightly colored minerals, including the bright green mineral found in the patina of the Statue of Liberty. For thousands of years, these visual cues have been the surest way to distinguish useful minerals, metals, and unwanted rocks.

The mining industry's exploration success rate (i.e., the number of large deposits discovered with an investment of $1) has been declining for decades. At KoBold, we sometimes talk about "Mining Anti-Moore's Law" (Eroom). As its inverted name suggests, it seems to be the opposite of Moore's Law. According to the anti-Moore law of mining, the number of deposits discovered per $1 invested has decreased by seven-eighths over the past 30 years. (The original anti-Moore law was aimed at a similar trend in the cost of developing new drugs.) ）

Geologically, the decline in newly discovered deposits is largely due to the fact that most of the easily discoverable minerals, such as surface deposits, have already been discovered. The new deposit will be buried deep underground, hidden by layers of rock.

In fact, most of the Earth's deposits are still waiting to be discovered. The chemical and physical processes that form these ores take place under conditions of temperature and pressure for several kilometers underground. That is, these deposits were not formed on the surface; Long after they were formed, tectonic processes brought small portions of them to the surface. This small part makes up the majority of the deposits mined today. The mining industry has the equipment and technology to mine deposits deep underground, and the problem is to find these deposits first.

You might think that the mining industry would invest heavily in exploration and R&D to improve exploration methods. But that's not the case. Over the past few decades, large companies have relied less and less on their own exploration projects and more on acquiring discoveries from other companies. Shareholders of mining companies expect dividends, not innovations.

At KoBold, we see exploration as a matter of information: discovering and analysing multiple types of data to reveal what we're looking for. In particular, in this information problem, getting more data types comes at a high price. Our solution is to combine AI systems with geoscience expertise to find out the information that best reduces uncertainty.

There is a wealth of geoscience information in the public domain, but it is fragmented and fragmented. Some of this information comes from government-funded geological surveys, and some from private companies that the government requires to make their findings public. The information is distributed across millions of datasets, including geological maps showing rock types observed at different locations, geochemical measurements of the concentrations of dozens of elements in rocks, soils, cores, plants, and groundwater samples, geophysical measurements of gravitational fields, magnetic fields, natural and induced currents, seismic waves, and nuclear decay radiation from heavy elements in the Earth's crust, satellite images measuring the spectral reflectance of minerals on the Earth's surface (including visible and infrared bands), and text reports describing field observations. All in all, the data is massive.

What's more, these datasets come from a wide range of sources, from state-of-the-art mass spectrometry measurements to century-old maps hand-drawn on linen. Each dataset has a purpose, and when combined in the right way, you get a powerful picture – if you can understand it.

Our data system, called TerraShed, parses information and transforms it into a standard form that can be accessed and searched by both humans and algorithms. Managing data and quality control it is only the first step. We then use a variety of algorithms to guide decisions about what data to collect at various stages of the exploration process, to understand whether a particular deposit is worth mining, and how the mine itself is being built.

TerraShed generates more than a simple "treasure map": it doesn't spit out "Marked Place X" based on data. Instead, we have hundreds of different proprietary modules that guide every decision in the exploration process.

Our exploration projects in the north of Quebec provide a good case study. First, we use machine learning to predict where nickel ore is most likely to be found in sufficient quantities to be worth mining. We use the underlying physical and geological data available in the region to train the model and supplement the model results with expert insights from geologists. In the province of Quebec, the models point us to a site less than 20 kilometers from the mine that is currently operating.

Once the relevant land rights were obtained, our geologists worked in a field camp on site, observing and surveying the outcropping rocks. On the more than 800 square kilometers of land we have specified, the options for rock sampling are almost limitless. However, this is not the case with time and money, and in our work area, the window time without snow on the ground is less than 3 months.

So the information challenge becomes: How do you decide which rocks to sample?

We built the Machine Prospector, which includes machine learning models and historical data such as information that has been previously discovered at other locations in the province. Prospectors help us predict which rocks should be sampled for a limited amount of time. Specifically, we are looking for deposits of magmatic sulphides rich in nickel and cobalt that have been formed over a long period of geological time.

With the predictions in hand, field geologists can begin to disperse their work. Someone travels to the site where such magmatic sulphides are most likely to be produced; Others went to the places where the uncertainty of the forecasts was highest. Collecting data at locations with high prediction uncertainty is more beneficial to improving the next generation of models than collecting data at locations identified by model predictions.

When the field team returns to camp in the evening, it uploads the day's data via satellite. Our data scientists work across the globe to retrain models based on new data points, resulting in new predictions, changing potential sample locations across the region and guiding teams in decisions about where to go next. By integrating new field information in near real-time, the model's adaptive prediction effectively shortens the learning period from one season to one day.

Compared to traditional forecasts using geological maps, our model produces 80% fewer false positives and false negatives for predictions. The process of constructing a geological map involves first observing rocks in fewer locations, and then using a set of rules and norms to extend the observations to a larger area. This means that traditional forecasting is largely an inference, and worse, it leads to uncertainty that cannot be quantified. In other words, we don't know how accurate such a map is. In contrast, KoBold's predictive model does quantify the uncertainty, which in turn guides our data collection efforts, as the rocks with the highest uncertainty tend to be the most valuable samples.

The results of an exploration project in northern Quebec in 2022 are a perfect example of the effectiveness of our unique approach to exploration.

Guided by the results of the AI system, our field team found a large area full of boulders, which geologist Lucie Mathieu believes is highly anomalous and not the typical igneous rocks that make up most of the boulders in the area.

Electromagnetic measurements showed that the boulders were unusually conductive, consistent with the mineral species we were looking for, and this result sparked our initial interest. We carry out daily time-domain electromagnetic measurements, which are collected by a helicopter towing a 30.5-metre-diameter transmitter coil. In the measurement, the transmitter emits a 7.5 Hz pulse current through a coil, which induces an induced current in the underground conductive material. At the end of the pulse transmission, the receiving coil detects the attenuation of the induced current in the subsurface, which allows us to build a three-dimensional model of the conductivity of the subsurface rock. The high conductivity of the ore we are looking for is just one of several factors that distinguish the ore from other rocks.

Helicopters and geophysical survey equipment are expensive, and in the north, the window for good weather is fleeting and unpredictable. Where to send the helicopter and how to make trade-offs between air coverage and spatial resolution are critical considerations.

We can use the collected data to build a 3D model of the possible location of the underground ore, which is a very difficult computational problem. Briefly, we have made limited measurements of the induced field of the two-dimensional plane of the earth's surface in an attempt to derive the properties of the three-dimensional volume of the subsurface (in this case, conductivity). There are countless subsurface rock formations here, consistent with the data on the ground.

The traditional practice in the industry is to build an optimal valuation model that fits a large number of parameters, which can easily exceed the number of data points. If you use n systems of equations to solve 2n unknowns, there is no unique solution to this problem. Traditional methods used in the industry choose one of many potential solutions, often containing assumptions that are inconsistent with geological processes and prone to confirmation bias.

To do better, we quantify the uncertainty of subsurface forecasts. Our machine learning models use far fewer parameters than traditional optimal valuation models, and the parameters are directly related to the key exploration question: How many conductors are present? How deep are they? How's it going? Do they have the same conductivity as those of high rich ores? The output of our model is a joint probability distribution of these parameters.

Ultimately, the availability of collected data depends on the ability to reduce uncertainty in the discovery of mineable deposits. Together with collaborators in Stanford University's Mineral-X program, we have developed a new method to quantify the usefulness of incremental data. In March 2022, we published what we call "information efficacy" in Natural Resources Research and used it to design drilling programs and other exploration programs in northern Quebec.

Over the course of the summer in Quebec, we drilled a total of 10 exploration holes, each of which was more than 1 kilometre away from the previous one. We combine the results of the predictive model with the expert judgment of our geologists to determine the location of each drilled well. The data collected from each hole indicates that we have found a conductive in a suitable geological setting underground, in other words, this could be a mineable deposit. Ultimately, we found nickel sulphide in 8 of the 10 exploration holes, which is almost 10 times the industry average at similar hole spacing.

We are also satisfied with the accuracy and specificity of the forecasts. For example, in hole KSC-22-004, our data scientists predicted that the conductor would be 130 to 170 meters underground. Drill down and we encountered highly conductive rock at 146m.

A few days before the end of the field season, we made a special discovery. This data helped to define the subsurface geological structure, so our team could determine the shape and size of the deposit with the most efficient drill holes in the next season, which will begin soon.

Assuming that our proven deposits in the region have a bright future as hoped, we can move on to the next deposit in search of another critical metal needed to electrify the planet. Overall, at least 1,000 new mines will need to be developed globally by mid-century to provide enough critical metals to produce enough electric vehicles to avoid the severe consequences of climate change. It's a daunting task, but by applying a new AI system like KoBold, we may be able to tap into new opportunities fast enough.

By Josh Goldman, Kurt House

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com