Skip to main content

ModEx Benchmark: AI for Benchmarking Model-Experiment Integration Workflows

EMSL Project ID


Recently, the U.S. Department of Energy (DOE), Office of Biological and Environmental Research (BER) launched and concluded the Artificial Intelligence for Earth System Predictability (AI4ESP) workshop. From this workshop, a critical conclusion that the DOE-BER community came to is the requirement to develop a new paradigm for Earth system predictability focused on enabling artificial intelligence (AI) across the field, lab, modeling, and analysis activities, called ModEx. Specifically, AI4ESP initiative wants to address the following key scientific challenge: How can we efficiently combine the multi-modal experimental/field data with multi-physics numerical models to accurately simulate water cycle and dramatically improve our understanding and representation of the Earth systems? Our proposed research focuses on a crucial aspect of the water cycle: Accurately estimating the 3D permeability field for modeling water flow and species transport in the subsurface. Accurate estimation of the permeability field is necessary for a predictive understanding of the infiltration process, which is a critical component of the water cycle. Due to the subsurface’s heterogeneous nature, much uncertainty is involved in estimating the 3D permeability field. Hence, capturing necessary structural features in permeability is essential for developing realistic subsurface process models for natural and engineered systems. There are various subsurface imaging techniques to collect data for estimating the 3D permeability field. Among such techniques, time-lapse electrical resistivity tomography (ERT) is a popular geophysical method to estimate 3D permeability fields from electrical potential difference measurements. Traditional inversion and data assimilation methods are used to ingest this ERT data into hydrogeophysical models to estimate permeability. Due to ill-posedness and the curse of dimensionality, existing inversion strategies provide poor estimates and low resolution of the 3D permeability field. Recent advances in deep learning provide powerful algorithms to overcome this challenge. In this project, we will develop a scalable AI framework and benchmark datasets to estimate the 3D subsurface permeability from time-lapse ERT. To test the feasibility of the proposed framework, we train AI inverse models on a massive multi-modal simulation dataset. This synthetic data represents the field study of stage-driven groundwater/river water interaction along the Columbia River, Washington, USA. The trained inverse models will be used to calibrate hydrogeophysics process models. Our approach is novel because once an AI model is trained, it can be re-used for inversion rapidly for the given system, with set physics and domain. Such a fast and accurate estimation of the permeability field may not be possible with traditional inversion (e.g., 104-105 forward model runs), which are very expensive. Our project is timely because we have acquired large-scale multi-modal field datasets needed for model-data integration. This field data consists of flow, temperature, and electrical conductivity time series collected at 20 well locations and 40466 time-lapse ERT measurements throughout six months at 352 electrodes. We will use EMSL’s high-performance computing resources to efficiently generate synthetic data using active learning and accurately train AI models. Our approach is innovative as this project will create open-source multi-modal and multi-physics benchmark datasets that do not exist within EMSL and BER communities to test AI-enabled ModEx workflows.

Project Details

Project type
Exploratory Research
Start Date
End Date


Principal Investigator

Maruti K. Mudunuru
Pacific Northwest National Laboratory

Team Members

Satish Karra
Environmental Molecular Sciences Laboratory

Glenn Hammond
Pacific Northwest National Laboratory