Research on Fluid Distribution Characterization Method of Tight Sandstone Reservoirs Based on Machine Learning Using Nuclear Magnetic Resonance Logging

Fluids in porous rocks can be divided into two categories according to their distribution patterns: irreducible fluids and movable fluids. Due to the influence of various geological processes, porous rocks exhibit a wide range of pore size distributions, leading to complex fluid distributions in pores of different sizes. The microscopic distributions of irreducible and movable fluids, namely the contents of irreducible and movable fluids in rock pores of varying sizes, can directly reflect the petrophysical properties of rocks, such as microscopic pore structure characteristics and seepage capacity. In the exploration and development of oil and gas resources, the quantitative characterization of the distributions of irreducible and movable fluids in reservoirs—especially the characterization of movable fluid distributions—is of great significance for reservoir evaluation, productivity prediction, and efficient reservoir development. However, owing to the limitations of actual core data, traditional modeling methods face bottlenecks at the data-driven level, posing challenges to the establishment of accurate fluid distribution characterization models. In this study, the fluid distribution laws in tight sandstones were first analyzed based on core experimental data. Then, the Generative Adversarial Networks (GAN) were used to expand the core dataset. The results of core data processing indicated that the fluid distribution laws of the generated data were consistent with those of the original data, which verified the effectiveness of the adopted data expansion method. Finally, the fluid distribution prediction model were established based on a Multilayer Perceptron (MLP) and realized the accurate characterization of the distributions of irreducible and movable fluids in tight sandstone reservoirs through core experiments and logging data processing.