New AI Model Could Improve Critical Mineral Discovery and Recovery

The key to finding more critical minerals may lie in understanding how fluids and chemicals move through rock. As these fluids travel underground, they can dissolve, transport, and concentrate valuable minerals in specific places.

Until now, studying these complex processes has required either time-consuming computer simulations or AI systems that need large amounts of training data. Even then, they sometimes produce unrealistic results.

Supported through an Environmental Molecular Sciences Laboratory (EMSL) user project, a team of researchers from the University of Houston and EMSL developed a new AI-powered tool that can predict how fluids and dissolved chemicals move through rocks and other porous materials underground. The new approach combines AI with the established laws of physics and chemistry to deliver faster, more reliable predictions.

"We developed a physics-informed machine learning framework that can rapidly forecast important parameters that cannot be measured using traditional experiments," said Kalyana Nakshatrala, University of Houston associate professor of engineering and principal investigator (PI) for the EMSL user project leading the effort.

The team's results were recently published in Transport in Porous Media.

Combining Physics and AI for Better Predictions

Modeling subsurface environments is particularly challenging because fluids and chemicals move through complex networks of pores and fractures, while researchers often have only limited data about what is happening underground.

The EMSL-University of Houston team's method uses a type of AI called a physics-informed neural network (PINN). Unlike conventional data-driven machine learning, which learns primarily from data, a PINN is trained using both data and the scientific laws that govern fluid flow and chemical reactions underground.

As the model learns, its predictions are continually checked against those laws of physics and chemistry. If it produces a prediction that violates those rules, it adjusts its calculations until the prediction better matches both the available data and the underlying science.

This approach offers several advantages over existing methods. Traditional computer simulations can accurately model underground processes but often require significant computing resources and can be difficult to apply when important information is missing or sparsely available. Conventional AI models can make predictions more quickly, but they typically require massive amounts of training data and may produce results that are not physically realistic.

By combining machine learning with established scientific principles, the team's framework can make faster predictions while remaining grounded in the physics and chemistry that control how minerals, fluids, and dissolved chemicals behave underground.

Reducing Inaccurate Results

A key part of the team's approach is ensuring that the model does not produce results that violate basic science.

For example, if a model predicts the amount of a mineral or chemical present in a rock formation, that value should never fall below zero. Yet small computational errors can sometimes produce these physically impossible results, especially when modeling critical minerals that occur in trace amounts and are difficult to detect accurately.

"In critical minerals, the concentrations are very, very small," said Maruti Mudunuru, an Earth scientist at Pacific Northwest National Laboratory and member of the project. "You're trying to find a needle in a haystack. When values are close to zero, even a small modeling error can lead to an unrealistic negative result."

The team's framework is designed to keep predictions within realistic bounds while still satisfying the physical and chemical rules that define the system. Nakshatrala said the challenge is balancing all of those requirements at once.

"If you try to enforce one aspect, some other aspect will be violated," Nakshatrala said. "In our framework, we try to manage all the errors of each chemical component to be very, very small."

The approach also gives researchers a way to estimate important factors that cannot be directly measured in experiments. In subsurface environments, scientists often do not have complete information about flow conditions, initial chemical distribution/abundance, reaction rates, or other parameters (e.g., permeability, porosity) that influence how minerals move and react. The new framework can help fill those gaps by combining limited data with existing scientific knowledge.

Advancing Critical Mineral Extraction

The new modeling capability has future potential for a variety of applications in critical mineral recovery efforts, including in situ mining, biomining, acid mine drainage, and recovery from waste piles or produced waters, Nakshatrala said. In these systems, researchers and industry need to understand how fluids should be injected, how long they should move through a material, and what conditions could maximize mineral recovery, he said.

By predicting how fluids, dissolved minerals, and chemical reactions interact underground, the framework can help researchers test fluid injection and recovery strategies virtually before they are evaluated in the field, potentially reducing costs and improving mineral yields.

"What should I pump? How fast should I pump? How long should I pump?" Nakshatrala said. "These are the kinds of questions you have to ask in order to make the project viable."

Laying the Foundation for Smarter Critical Mineral Recovery

Nakshatrala said the team's framework marks an important starting point in the development of physics-informed machine learning tools for critical mineral recovery.

So far, the framework has been tested on simplified problems that capture key processes involved in critical mineral recovery. Future versions, he said, could incorporate additional chemistry, biology, or microbial processes as researchers work toward more realistic models of underground systems.

Mudunuru said the work connects to EMSL's broader efforts to link modeling, experiments, and data. Through EMSL Community Science Campaigns, researchers are testing novel approaches that use experimental platforms such as laboratory chips, structure and chemical imaging data, and porous materials to better understand how minerals move and react. These experiments can help validate models, while the models can help guide future experiments, he said.

Looking further ahead, Nakshatrala said the long-term goal is to move toward AI-enabled digital twins for critical mineral and leaching operations. A digital twin is a computer model that can update itself as new data become available, creating a feedback loop between experiments, field observations, and predictions. While this study is only an early step in that direction, it provides a foundation for adaptive tools that could eventually help guide critical mineral recovery in real time, he said.