This project develops a physics-informed machine learning (PIML) methodology to upscale soil structural information (e.g., X-ray and microscopy images) to continuum-scale process models (e.g., DOE codes such as PFLOTRAN). The proposed methodology uses recent advances in machine learning (ML), such as physics-informed neural networks (PINNs) and variational auto-encoders (VAE), and pore-scale modeling using the Lattice Boltzmann method (LBM). The methodology will result in an ML upscaler product to estimate a soil sample’s permeability by considering the 3D soil structural information and ML-emulated fluid flow that approximates the solution of the Navier-Stokes equation. The ML upscaler can eventually be used in continuum-scale flow codes such as PFLOTRAN or ATS for further higher-scale simulations. The lead PI is Prof. Kalyana Nakshatrala, who has extensive expertise in pore-scale modeling and numerical simulations. To successfully execute this project, EMSL’s expertise in AI/ML (e.g., neural networks), network science (e.g., flow and transport in graphs), and image segmentation (e.g., Fiji and XCT data from MONet) will be leveraged. Building on EMSL’s modeling and AI/ML expertise will enable us to develop workflows for the BER community. Successful execution of our research requires EMSL’s HPC resources, soil sample data (e.g., labeled XCT images from MONet), and technical expertise on AI-enabled ModEx, Fiji AI segmentation software, and pore-scale modeling. We anticipate the usage of ~200,000 CPU node-hours and 5TB storage, estimated based on our previous works.