Computing, Analytics, and Modeling
New AI and Machine Learning Tool Predicts Subsurface Processes Important for Critical Mineral Extraction
A multi-institutional team combined a physics-informed machine learning approach with established scientific models and data and validated that it accurately predicts the spatial distribution of chemical reactions involved in subsurface resource recovery.
Reliable prediction of subsurface chemical processes is important for improving strategies for critical mineral recovery and other energy resource recovery applications from the subsurface. (Graphic by imaginima, iStock)
What is the science?
Chemical reactions that occur underground influence how critical minerals move, concentrate, and become available for extraction. However, these reactions often take place within complex geologic environments where fluid flow, chemical transport, and chemical transformations occur simultaneously, making them difficult to predict accurately.
A multi-institutional team of researchers developed a physics-informed neural network (PINN) framework, which uses machine learning (ML) to combine established scientific knowledge, existing reactive transport models, and the laws of physics governing these processes within a single framework. By integrating these sources of information, data, and models and then validating the PINN-based model outputs, the team demonstrated that ML could be used to successfully reproduce complex subsurface chemical behavior while requiring less computational effort than that used by conventional simulation methods. The findings demonstrate a new way to model underground chemical systems with both speed and accuracy.
What is the impact of this scientific work?
Reliable prediction of subsurface chemical processes is important for improving strategies for critical mineral recovery and other energy resource recovery applications from the subsurface. Many existing modeling approaches require substantial computing resources, while purely data-driven AI methods often require large amounts of training data. By combining physics-based understanding and existing computational models, the PINN-based framework provides a more efficient way to evaluate subsurface behavior. The capability could accelerate the assessment of resource recovery strategies, support biotechnology and energy-related research, and enable scientists to investigate problems that would otherwise be impractical to study using conventional simulations alone.
What is the summary of this scientific work?
Accurately predicting how chemicals move and react underground is essential for applications such as critical mineral extraction, subsurface energy resource production, and understanding current and future biogeochemical reactions and nutrient cycling. However, conventional simulations can be computationally expensive and difficult to apply to complex subsurface systems. To address this challenge, a multi-institutional team of researchers led by the University of Houston developed a PINN, a type of ML model that incorporates known laws of physics directly into the model. The research was conducted as part of a Large-Scale Research project through the Environmental Molecular Sciences Laboratory (EMSL), a Department of Energy Office of Science user facility. Rather than relying solely on training data, the team integrated scientific knowledge from established reactive transport models into the PINN framework, allowing the model to learn from both physical principles and computational simulations. The team then trained and tested the framework using standard reactive transport test cases representing fluid flow, chemical transport, and chemical reactions in subsurface environments. EMSL supported the research through computational resources and expertise. After comparing the model's results with conventional simulations, the team found that the PINN-based framework accurately captured chemical mixing, sharp reaction zones, and complex reaction behavior while using fewer computational resources. These results demonstrate that physics-informed ML can combine the strengths of AI and traditional scientific modeling to improve predictions of subsurface processes, providing a foundation for future applications in critical mineral recovery, subsurface energy, and related fields.
Who are the contacts for this scientific work?
- Kalyana B. Nakshatrala, University of Houston
knakshatrala@uh.edu
- Maruti Mudunuru, Pacific Northwest National Laboratory and Environmental Molecular Sciences Laboratory
maruti@pnnl.gov
How was this scientific work funded?
This research was conducted as part of a Large-Scale Research project through the Environmental Molecular Sciences Laboratory (EMSL), a Department of Energy Office of Science user facility sponsored by the Biological and Environmental Research program. EMSL provided computational resources and expertise that supported the development and evaluation of the physics-informed machine learning framework. Additional support was provided through the University of Houston Additive Manufacturing Institute and the University of Houston-Chevron-Energy Fellowship.
What is the publication citation?
K. Adhikari, et al. "Reactive Transport Modeling with Physics-Informed Machine Learning for Critical Minerals Applications." Transport in Porous Media 157, 91 (2026). [DOI: 10.1007/s11242-026-02301-9].