Indaba: How AI in mining is like self-driving cars

The planet’s most advanced artificial intelligence isn’t a chatbot, it’s a robotic vehicle, and Stanford University researchers say mining should pay attention.

A team at Stanford’s Mineral-X initiative, led by Professor Jef Caers, is applying an intelligent-agent approach to exploration, mine planning and mineral processing. It is building systems that plan actions while acknowledging uncertainty rather than assuming a single fixed outcome. The models are designed to update decisions as new data arrive, mirroring how driver-less autos recalculate routes in real time.

“The most sophisticated AI today in the world is in San Francisco, and it’s called a self-driving car,” Caers told delegates at the Mining Indaba. “It faces changes in the road or conditions that were unexpected. Sounds familiar to mining? Well, it is.”

The academic’s work adds to the explosive growth of AI applications that the mining world is only starting to grasp. Industry analysts tracking digital transformation such as CostMine Intelligence report that predictive maintenance and real-time process optimization can cut unplanned downtime by 20% to 40% and lift plant throughput by mid-single to low-double digits.

For a large copper or gold operation, that can translate into tens of millions of dollars in annual value through higher recoveries, lower energy intensity and fewer disruptions.

New eyes

In Brazil’s Carajás region, an undisclosed copper company identified hundreds of magnetic anomalies that resembled a known iron oxide copper-gold deposit, yet the differences between promising and marginal targets were subtle.

“There was a very simple question, but a very expensive one, where do we drill?” Stanford PhD researcher Sofia Mantilla Salas told the session on Wednesday. “The problem is that the targets that they had look very similar to the human eye.”

Instead of ranking anomalies visually, her team trained AI to identify the deposit’s signature across multiple geophysical datasets and then scan broad areas automatically for similar patterns. The system can detect combinations of magnetic, radiometric and structural signals that would be difficult to interpret consistently by eye, narrowing the list of drill candidates and reducing false positives before capital is committed.

“It’s not that the AI replaces the geologist,” she said. “I am a geologist, but it actually helps us to do our job better.”

Drilling planner

Caers noted that the framework has already been used in Zambia for sequential drill planning. The model evaluates where to drill next while explicitly accounting for what remains unknown, rather than relying on a static interpretation frozen at the start of a campaign.

In mine planning, the research focuses on uncertainty that can undermine project economics if left unmodelled. Gloria Quispe Oruro, a mining engineer and PhD researcher, is applying probabilistic simulations to groundwater behaviour in copper deposits. Conventional plans often assume that subsurface conditions remain stable once defined, she said, yet water flow and pressure responses can alter sequencing, slope stability and capital timing.

“We think that the subsurface is going to stay the same way it was when we explored,” she said.

Her approach runs thousands of simulations to test variations in permeability and porosity, estimating how dewatering strategies influence development over a decade or more. By quantifying risk in advance, operators can adjust schedules, infrastructure design and capital deployment with clearer visibility into potential constraints.

Processing

Processing presents another layer of uncertainty. Feed composition can vary daily, affecting concentrate grade and recovery.

“Even if you know the feedstock, it’s quite difficult to predict what comes out to concentrate grade and recovery,” said William Zu, a materials scientist working with Mineral-X.

The team is exploring adaptive blending and flotation design tools that respond to changing ore characteristics. Rather than replacing engineers, the systems evaluate multiple operating scenarios simultaneously and recommend adjustments under fluctuating conditions.

“Humans are not the best at making decisions under uncertainty,” Zu said, adding that AI can “augment human decision making in these spaces.”

Data quality

Across exploration, planning and processing, researchers repeatedly returned to data quality as the limiting factor. Mining companies already collect vast quantities of geological and operational information, but inconsistent formatting and siloed storage reduce its usefulness.

“In the mining industry, there’s a wealth of data that’s available now that can really feed into these AI algorithms,” Zu said. “But definitely the challenge to that is data quality.”

Quispe Oruro framed the issue in financial terms.

“Instead of using the data as something that is sitting on a hard drive, they are using that data as an investment. So when you invest in that data, you also want returns on it,” she said.

Caers was direct about the boundaries of the technology.

“Is AI going to make a discovery? I don’t think that’s going to happen,” he said. “I think people are going to use data better, and that is going to lead to successes.”

“Let’s just make better decisions and make better use of data.”