Machine learning-based nuclear magnetic resonance measurements of hydraulic properties in hydrate-bearing sediments

Abstract

Permeability and water retention curves of hydrate-bearing sediments play pivotal roles in comprehending the dynamics of fluids within geologic hydrate systems and have direct impacts on gas production efficiency. Previous investigations into these hydraulic properties have been constrained to cursory field estimates or laboratory assessments of localized small samples. This hinders high-fidelity numerical simulations with regard to the evolution of widely distributed gas hydrate in subsurface and the evaluation of its potential as an energy resource. This study represents the inaugural comprehensive compilation of permeability data measured in global hydrate survey areas, aiming to develop a data-driven characterization of hydraulic properties from nuclear magnetic resonance (NMR) measurements. Interpretable machine learning substantially improves the conventional Schlumberger-Doll Research (SDR) equation, establishing correlations with clay content, porosity, formation factor, and hydrate saturation to accommodate diverse lithologies across regions. By combining Kozeny's theory on the nature of the pore system with the data-driven SDR equation, NMR data can be used to efficiently furnish precise and reliable assessments of permeability and water retention curves for hydrate-bearing sediments, prior to extensive specialized core measurements. The results show that our model achieves an overall predictive accuracy for permeability of R² = 0.924 and MAE = 0.319 across typical hydrate occurrence zones, with 97.6% of samples displaying absolute errors within one order of magnitude.