Application of Machine Learning Tool to Separate Overlapping Fluid Components on NMR T2 Distributions: Case Studies from Laboratory Displacement Experiment and Well Logs

Abstract

Fluid typing, meaning fluid identification and quantification of each phase, is a significant challenge in NMR data postprocessing, particularly when the fluid spectral distributions overlap in one-dimension (T1, T2 spectra) or two dimensions (T1T2 or DT2 maps). Typical examples are extra-heavy oil and clay-bound water (CBW), heavy oil and capillary-bound water (BVI), free water and light oil or light oil-water and oil-base mud filtrate (OBMF). In these cases, technical limitations in data acquisitions and constraints in the inversion algorithms result in poor spectral resolution for those fluids with very similar physical-chemical properties. This makes very difficult the interpretation of NMR measurements from the laboratory as well downhole.

We present two case studies: one focused on determining water saturation (Sw) in core samples in a water-oil displacement experiment in the laboratory; the second is about determining the permeability by identifying OBMF from an NMR well log in a medium to light oil-bearing formation. In both cases the targeted fluid component was determined using blind source separation based on independent component analysis (BSS-ICA), which is a machine learning tool capable of separating spectral T2 components (sources) given their statistical independency.

The results from the displacement experimental show a high correlation (R2 higher than .85) between saturation from the BSS-ICA derived water component and the estimated value from known injected water volumes. In the well log case, the results show that the presence of OBMF and its volume are a good indicator of rock quality of the reservoir layers, as confirmed from several core measurements. Beyond this, the OBMF obtained from BSS-ICA decomposition is used as a key variable in a newly proposed permeability equation. After core calibration the OBMF-based permeability was found to be more representative than the permeability derived from the widely used Timur-Coates equation.