Novel methodology for accurate resolution of fluid signatures from multi-dimensional NMR well-logging measurements

A novel methodology for accurate fluid characterization from multi-dimensional nuclear magnetic resonance (NMR) well-logging measurements is introduced. This methodology overcomes a fundamental challenge of poor resolution of features in multi-dimensional NMR distributions due to low signal-to-noise ratio (SNR) of well-logging measurements. Based on an unsupervised machine-learning concept of blind source separation, the methodology resolves fluid responses from simultaneous analysis of large quantities of well-logging data. The multi-dimensional NMR distributions from a well log are arranged in a database matrix that is expressed as the product of two non-negative matrices. The first matrix contains the unique fluid signatures, and the second matrix contains the relative contributions of the signatures for each measurement sample. No a priori information or subjective assumptions about the underlying features in the data are required. Furthermore, the dimensionality of the data is reduced by several orders of magnitude, which greatly simplifies the visualization and interpretation of the fluid signatures. Compared to traditional methods of NMR fluid characterization which only use the information content of a single measurement, the new methodology uses the orders-of-magnitude higher information content of the entire well log. Simulations show that the methodology can resolve accurate fluid responses in challenging SNR conditions. The application of the methodology to well-logging data from a heavy oil reservoir shows that individual fluid signatures of heavy oil, water associated with clays and water in interstitial pores can be accurately obtained.