Utilizing Deep Learning to Simulate Transient Fluid Pathways: Integrating 4D Synchrotron Tomography Experiments with Natural Observations

Abstract

Determining the dynamics of fluid-rock interactions is key to deciphering processes in the lithosphere and their importance in industrial applications, including the storage of CO2 and hydrogen, as well as in the development of geothermal energy technologies. Many geological occurrences show that when fluids interact with rocks, they can form their own pathways by developing a temporary network of connected pores as a result of mineral replacement reactions. Nonetheless, following the reaction's cessation, most pores become isolated retarding fluid flow. Here, we investigate transient porosity generation phenomena by conducting 4D (3D plus time) synchrotron tomography experiments using a salt analogue system. To capture the rapid evolution of reaction-induced pore networks, we acquire a full tomography volume every minute. After segmenting this extensive dataset using a deep convolutional neural network, the dynamics of the reaction are quantified by spatiotemporal correlation functions. These high-order statistics enable us to evaluate the evolution of the structural and morphological properties of the induced pore network during the salt replacement reaction. Inspired by image editing techniques in computer vision, we then train a leading-edge generative model, StyleGAN2-ADA (Karras et al., 2020[P(1] ), utilizing the 4D tomography dataset. Our results demonstrate that these advanced generative models can accurately simulate the microstructural evolution observed in the experiment. Next, we delve into the potential of deploying the trained model to reconnect isolated pores observed in data sets of natural mineral replacement reactions. We finally apply a voxel-based finite element method to simulate fluid flow in the salt and natural system. Our deep-learning method not only pioneers in determining transient fluid-rock interaction phenomena, but also, for the first time, enables a direct estimation of reaction-induced permeabilities. Putnis, A. Mineral replacement reactions. Rev. mineralogy geochemistry 70, 87–124 (2009). Karras, T. et al. Training generative adversarial networks with limited data. Adv. Neural Inf. Process. Syst. 33, 12104–12114 (2020).