Learning Multiscale Subsurface Flow Physics for the Energy Transition

Subsurface geological formations are critical to the energy transition, offering both carbon storage and the potential for terra-watt-scale energy storage. While numerical modeling has traditionally been the foundation of subsurface management, machine learning (ML) presents a compelling alternative for modeling complex subsurface flow dynamics with unprecedented speed and accuracy. This seminar explores two novel ML frameworks operating at different scales.

At the reservoir scale, I will present the Adaptive Physics Transformer (APT) for modeling Aquifer Thermal Energy Storage (ATES) systems, a promising technology suite for decarbonizing the heating and cooling sector. APT learns reservoir temperature and pressure distributions across irregular grids and adaptive meshes. Its flexible architecture allows the training process to adaptively focus on heat and cold plumes to ensure efficient and accurate predictions.

At the pore scale, I will introduce the Pore-scale Graph Network Simulator (Pore-scale GNS), which learns multiphase flow dynamics directly from 4D micro-CT experiments with particle tracking. This multi-modality architecture concurrently predicts menisci movement and velocimetry information. The trained model can extrapolate to unseen experimental conditions, providing compelling evidence for a new framework of ML-based pore-scale physics modeling to bridge the gap between pore and field scales.

Speaker: Gege Wen, Imperial College, London, UK