Efficient, precise mining is more of an art than a science. Mining operators need more accurate estimations of where target minerals are located beneath the ground so they can drill and extract once instead of multiple times. Stratum AI, founded by Farzi Yusufali and Daniel Mogilny in April 2020, provides a deep learning algorithm for mine operators to use to better model the exact location of where valuable minerals are located underground. The startup resides in Toronto.
Stratum AI previously raised capital in September 2020 and used said funds to work with clients and continue improving their technology. There are 12 employees working for Stratum AI. The startup’s competitors are mostly incumbent, speciality mining consultant groups.
Frederick Daso: What are the traditional ways mine owners and operators conduct geological surveys to assess the mineral content of the surrounding land around and beneath them?
Farzi Yusufali: Under the guidance of geologists, mine owners and operators drill out long cores of rock for analysis, then assay each sample to determine the content. This is done repeatedly using a predetermined spacing/orientation decided by each company. Geologists use that drill data, their experience, and industry-standard geostatistics to “fill in the blanks” between the drill holes to create a 3D map of the mineral content in the ground.
Stratum uses that same data to create a more comprehensive 3D map that better represents the mineral content in the ground. We do that by using our own algorithms to “fill in the blanks,” and we don’t need to do our own data collection.
Daso: What features of the mined data set Stratum receives make it suitable for deep learning models to help tackle in the form of a solution? Is the data structured or unstructured as you ingest it into the data pipeline?
Yusufali: Drillholes (the samples taken from drilling) and production data don’t generally differ greatly from one mine to the next. In essence, you have four columns: x,y, and z coordinates and a number denoting the amount of a mineral in the sample.
The challenge is that a mining dataset, like the one described above, is not naturally compatible with machine learning—or deep learning, for that matter. Machine learning (and deep learning) are well-suited for images and text; mining data is collected less extensively (covering at most 10% of the total area), has non-trivial quality issues, and is not generally collected systematically. But the proprietary Deep Learning system at the core of Stratum’s IP can analyze the limited data and come up with a more accurate representation of what is actually in the ground.
Daso: What segment is Stratum choosing to tackle within the trillion dollar market for mined minerals first, and why?
Yusufali: Stratum’s technology is commodity-agnostic. Our first clients were copper and gold miners, representing the two major types of commodities, bulk and precious; therefore, we can presumably model any mineral or other information useful to the mine, like ore type. Working at complex mines with acute issues allows us to use our 3D maps to affect immediate changes to how the mine operates. In a sense, our DL models are very versatile because we can create maps of anything the mine extracts as long as the data is collected the same way as other drillhole and production data.
As far as markets are concerned, we do well whether overall global economic prospects are rising or falling. When economic expectations weaken, we get increased interest from precious-metal miners, and when the economy is rising, we get more interest from base metal miners. But no matter which market prices are heading, it always makes long-term economic sense for miners to invest now to test the technology for themselves and fix their problems. That way, they will be better positioned to weather the storm when their market weakens.
Daso: How does Stratum enable cost-efficient extraction of these minerals after ostensibly predicting their presence and approximate location underground?
Yusufali: Accurate 3D maps of the mineral distribution underground affect extraction in two ways: 1) reducing waste and 2) increasing yield.
In terms of waste reduction, mines often struggle with defining the boundaries between economic and waste material. Stratum’s 3D maps provide that resolution, characterizing each unit volume of rock as either ore or waste. This allows operators to avoid waste as much as possible by reducing the cost of extracting material that has no value.
In terms of boosting yields, mines often miss economic clusters of minerals underground, especially in more complex assets. Stratum’s 3D maps are particularly adept at identifying areas of high-grade material, even when far away from measured data. The value of finding missed material is pretty self-explanatory!
Daso: Are there long-term plans for Stratum to develop autonomous or teleoperated mining vehicles with the startup’s proprietary technology to enable “fully AI-operated mines” within the coming decades?
Yusufali: Stratum’s vision is to become the AI brain of a mining operation; tangibly, that means enabling mines to make both micro and macro decisions dynamically to maximize a mine’s profit margins. To do that, Stratum’s technology will need data collected by sensors, vehicles, and equipment regularly. Stratum will integrate with the best of the vehicles (autonomous or otherwise) as time goes on; if there isn’t hardware to collect the data we need, Stratum will take on that role and develop it internally.
Daso: What is your team’s philosophy on how humans and AI will interact in these “futuristic AI-operated mines?”
Yusufali: We’re far from generalized AI running every part of a mining operation. Our view is that, at least for the next couple of decades, human-AI interaction will need. AI will process many types of mine data quickly to give an overview of the mine’s performance and suggest actions when performance isn’t optimal. Humans will add data and constraints related to the operation and markets, the AI will “crunch the numbers” to produce a new prediction, and humans will then make decisions based on the AI’s output.