38-2 Advances in progressive transfer learning for subsurface storage and energy recovery systems

Session: Geoscience and Hydrogeology in the AI Era: From Predictive Models to Real-Time Applications

Presenting Author:

Author:

Abstract:

Recently artificial intelligence/machine learning (AI/ML) has been increasingly developed for various subsurface energy activities. Developing a real-time, accurate forecasting model that accounts for geological uncertainties and operational limits is crucial to optimizing subsurface operations and monitoring. AI/ML shows great promise in optimizing operations (e.g., injection and extraction through wells) and managing associated challenges. However, ML-driven subsurface modeling is often limited by the need for extensive datasets—usually produced by high-fidelity simulators. To address these constraints, we use a progressive reduced-order modeling framework [1] that reduces data requirements, making ML-based subsurface modeling feasible. Our approach focuses on three main goals including (1) transfer learning to apply insights learned from pre-trained models to new physical systems, (2) reduction of training data requirements through efficient knowledge transfer, and (3) model flexibility to handle multiphysics challenges from single phase flow to coupled multiphase flow-poroelasticity. Our framework's reliability and applicability are demonstrated through its validation with simulation data from the Illinois Basin–Decatur Project (IBDP) where a million tonnes of CO2 has been injected. By incorporating iterative L1 pruning, we ensure that only essential information is transferred, avoiding over-parameterization. We will demonstrate that progressive learning framework with improved neural operator model(s) (p-INO) can significantly enhance model generalization and convergence, especially in data-scarce scenarios, outperforming baseline ML models across all training conditions. In addition, a low rank adaptation (LoRA) for the INO can reduce model complexity with a minimal loss in accuracy, offering an efficient alternative when computational resources are limited. Overall, combining progressive learning with parameter-efficient fine-tuning techniques like the LoRA provides a robust framework for scalable and accurate modeling. Future work may explore adaptive compression strategies to optimize this balance further. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-8250

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.