SRI’s machine learning-based geospatial analytics platform, already adopted by the USGS, is poised to make waves in the mining industry.
Today, most mineral discovery starts in front of a computer rather than in a remote canyon. We now have vast amounts of digitized information — geological maps, mineral assays, geochemistry samples, remote sensing data, and much more — that help the mining industry zero in on the best potential locations for future mine sites.
That said, information collected over many decades exists in numerous incompatible formats. This creates challenges for United States Geological Survey (USGS) scientists tasked with manually synthesizing historical maps and other data to inform predictions about mineral deposits. The dynamic also creates barriers for mining companies that want to leverage USGS data and their own proprietary data to prospect for new mine sites in the United States.
A few dozen critical minerals — like lithium, graphite, and ytterbium — drive advanced IT and defense technologies. Countries with efficient access to those minerals will have an economic and security advantage.
Why is this such an urgent problem? Mostly because, amid intensifying global competition to locate and exploit rare earth elements, there’s a pressing need for the United States to quickly turn existing data into new sources of rare earths and other essential minerals.
Recognizing the need for new domestic sources of critical minerals, SRI researchers are developing a machine learning-driven approach that fast-tracks mineral discovery.
Minerals for national security and industrial advantage
SRI’s work on mineral discovery began when the USGS and the Defense Advanced Research Projects Agency (DARPA) tapped SRI for a research effort called CriticalMAAS, which focuses on the USGS critical mineral assessment process.
“The USGS’s standard critical mineral assessment workflow can take up to two years for a single mineral commodity,” observes Hang-Pang Chiu, a principal investigator for SRI’s CriticalMAAS work. “The goal here was to build an automated AI-assisted capability that could really speed up this process.”
This is no mere academic exercise: A few dozen critical minerals — like lithium, graphite, and ytterbium — drive advanced IT and defense technologies. Countries with efficient access to those minerals will have an economic and security advantage. In the United States, government initiatives are already transforming domestic mining markets, and commercial actors see increasing opportunities to ramp up domestic mining activity.
The initial impact of the CriticalMAAS program came into focus earlier this year, when a hackathon demonstrated that AI tools can compress the USGS’s critical mineral assessment timelines from two years into a mere two-and-a-half days.
Clear data, better predictions
Working on the CriticalMAAS program, teams in SRI’s Center for Vision Technologies recently advanced two key aspects of AI-infused mineral deposit discovery: Extracting relevant data from pre-existing geological documents and then building an AI-based system that predicts the most likely locations of specific mineral deposits.
SRI’s work on knowledge extraction, completed in collaboration with Craig Knoblock of USC’s Information Sciences Institute, focused on gleaning usable data from a wide array of heterogeneous USGS documents. Using large language models, the team parsed documents (including historical mining reports), created an automated and standardized approach to location data labelling and mineral deposit type classification, and organized the data into layers that enabled flexible search queries from USGS geologists.
“The terms that geologists use have shifted over time, making it difficult to reconcile historical records. LLMs and embedding models can handle these variations in semantic meaning.” — Meng Ye
“The terms that geologists use have shifted over time, making it difficult to reconcile historical records,” explains Meng Ye, the principal investigator for SRI’s data extraction work. “LLMs and embedding models can handle these variations in semantic meaning.”
To architect the platform’s predictive mapping capabilities, SRI developed a machine learning approach that relies heavily on self-supervised learning. The biggest challenge in predicting new mineral sites comes down to data label limitations. For some of the most valuable minerals, only a handful of sites in the United States might be known. And across the board, much of the relevant data is not labeled, making it difficult for AI systems to exploit. Self-supervised learning enables the model to learn, despite a lack of well-labelled information, the generic features of areas amenable to mineral exploration.
“These generic features can be used, with some fine-tuning, to produce mineral prospectivity maps,” explains Angel Daruna, an advanced computer scientist at SRI.
A powerful new AI tool for the mining industry
Now that USGS geologists are actively using SRI’s platform for critical mineral assessments, SRI is aiming to bring this capability to bear on other mining industry challenges where multimodal geospatial data can be used to facilitate faster decision making.
For example, the platform can de-risk mineral exploration and replace costly on-site work with off-site analytics. One significant cost in mineral discovery is exploratory drilling. Exploratory drill holes are essential to finding and developing new mines, yet siting these drill holes can be a guessing game. Leveraging SRI’s geospatial reasoning approach, mining companies could integrate publicly available data with their own proprietary logs to reduce exploratory drilling.
Moreover, the “explainability” inherent in SRI’s approach is designed to reduce risk in mineral exploration. Precise, quantifiable information about the model’s certainty or uncertainty would allow boards and regulators make confident data-driven decisions. In the future, the platform may also support feasibility studies (design, permitting, and environmental impact), incorporate regional reconnaissance (airborne geophysics and satellite imagery), and analyze factors like land status and tectonic belts.
“We see applications beyond mining as well,” Chiu says, “such as precision agriculture and renewable energy siting — really, any industrial workflow where geophysical realities are a deciding factor.”
To learn more about SRI’s work on geospatial reasoning, contact us.
This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR00112390130.
