USC Engineers Develop AI Model for Subsurface Fluid and Heat Flow Predictions

Image Credit: Gloria Liu | Splash

Researchers at the University of Southern California's Viterbi School of Engineering are developing an artificial intelligence model to predict subsurface flows of water, carbon dioxide and heat, with applications in geologic carbon storage and geothermal energy recovery.

The project, titled "CAIG: Advancing Subsurface Flow and Transport Modeling with Physics-Informed Causal Deep Learning Models" and known as PINCER, combines machine learning with physical principles to address complex interactions in underground systems. It is led by Yan Liu, a professor in the Thomas Lord Department of Computer Science, electrical and computer engineering, and biomedical sciences, as principal investigator, and Behnam Jafarpour, a professor of chemical engineering and materials science, electrical and computer engineering, and civil and environmental engineering, as co-principal investigator. The effort began on September 15, 2024, and is set to continue through August 31, 2027.

Project Background and Funding

PINCER tackles challenges in subsurface modeling, where systems involve interactions among rocks, fractures and fluids, complicated by uncertain and heterogeneous rock properties across scales. Traditional simulations often use uncertain inputs and simplified physics, while data-driven AI requires large datasets that are typically scarce for underground environments.

The initiative integrates advances in AI for processing diverse multimodal datasets with domain knowledge, embedding physical principles like fluid flow equations into deep learning frameworks to enhance accuracy and reduce data needs. Seed funding came from USC's Ershaghi Center for Energy Transition, facilitating initial collaboration between Liu and Jafarpour.

Liu's work includes machine learning for time series, explainable AI and physics-informed methods in climate and health domains. Jafarpour, holder of the Energi Simulation Industrial Research Chair in Subsurface Energy Data Science since July 2022, focuses on data assimilation, optimization and inverse modeling for reservoir engineering.

The project is funded by the U.S. National Science Foundation's Collaborations in Artificial Intelligence and Geosciences program, with an award of US$599,999 announced in August 2024 as part of a US$20 million effort to support interdisciplinary AI-geoscience teams. This program promotes mutual advancements, refining AI techniques through geoscience problems while improving earth system predictions amid climate pressures.

Methodology and Technical Approach

PINCER develops novel deep learning architectures that adhere to the structure of fluid flow equations, accounting for uncertainties and adapting based on incomplete multiphysics monitoring data. It incorporates causal inference to capture spatial and temporal relations in flow and transport dynamics, shifting from purely data-driven or model-based methods to a hybrid framework.

The approach includes flexible deep learning for inferring heterogeneous rock properties from sparse observations, aiming to predict subsurface behaviors more reliably. Jafarpour has described this integration as leveraging "the strengths of both fields" to enhance physical consistency. Liu has noted its potential for breakthroughs in "grand challenges" like accurate CO2 storage predictions.

Applications in Energy and Climate Contexts

PINCER targets geologic carbon storage by improving predictions of CO2 dynamics, aiding reservoir identification and response forecasting to support carbon capture efforts. Related research from the team has explored physics-guided deep learning for CO2 migration.

For geothermal energy, the model focuses on recovery processes, contributing to sustainable resource management. It also applies to groundwater aquifers, enhancing overall environmental sustainability amid climate change.

These areas align with U.S. priorities for energy transitions, including tax incentives under the Inflation Reduction Act that support carbon storage and geothermal development.

Impacts and Challenges

The project emphasizes societal benefits, such as better management of earth resources, while training underrepresented students in AI-geoscience integration. Challenges include handling non-linear behaviors and ensuring model adaptability across varied subsurface conditions, as highlighted in reviews of physics-informed neural networks.

PINCER aims to advance hybrid AI models for subsurface predictions, potentially influencing geoscience tools by project end in 2027. This reflects growing trends in AI-geoscience collaborations, with analyses suggesting increased adoption in energy sectors over the next decade, provided interpretability and scalability issues are addressed.

License This Article

Source: US National Science Foundation, USC Viterbi

$5.00

$10.00

$20.00

$30.00

Custom Amount

Please enter an amount

$

One-Time Donation Monthly Donation Quarterly Donation Annual Donation

3% Cover the Fee

Support us to keep delivering insightful content on AI development. Your contribution fuels our mission to inspire and inform! Support us to keep delivering insightful content on AI development. Your contribution fuels our mission to inspire and inform!