Learning and Inferring Multiphase Flow Dynamics in Porous Media using Scientific Machine Learning: Application to the "FluidFlower" CO2 Injection Experiment

Abstract

Abstract

Accurate prediction and parameter identification of multiphase flow in porous media remain central challenges in geological carbon dioxide storage due to strong nonlinearities, high-dimensional parameter spaces, and limited observational data. We present a machine learning framework that integrates surrogate modeling and Bayesian inference to enable efficient forward prediction and inverse parameter estimation for CO2-brine flows in geological media. The approach is demonstrated using the "FluidFlower" experimental rig, a controlled laboratory system that provides high-resolution, time-resolved observations of CO2 migration in heterogeneous porous media. A convolutional neural network surrogate is trained on high-fidelity numerical simulations to learn the evolution of CO2 saturation and dissolved CO2 concentration fields over a wide range of multiphase flow properties. The trained surrogate is embedded within a Markov chain Monte Carlo framework for parameter inference conditioned on experimental observations. Results show that the surrogate accurately captures large-scale CO2 plume migration, dissolution dynamics, and multiphase flow behavior while providing orders-of-magnitude acceleration compared to traditional simulations. Embedding the surrogate within a Bayesian framework enables computationally tractable exploration of the parameter space and reveals both identifiable and non-identifiable parameter combinations that produce similar plume behavior. By leveraging spatially and temporally resolved full-field observations, the framework substantially improves agreement between simulations and experiments compared to previous manual calibrations based on limited plume-scale metrics. Analysis using progressively increasing observation horizons further shows that observations become more informative once the plume interacts with geological features such as faults and sealing layers.