Improved Quantification of Nuclear Magnetic Resonance Relaxometry Data via Partial Least Squares Analysis

Abstract

Nuclear magnetic resonance relaxometry measurements are frequently used to quantify sample constituents. The standard approach for quantification involves converting the time-domain data to a distribution of characteristic times, either by fitting a fixed number of exponentials or performing an inverse Laplace transform, and then integrating the area under the peaks. We evaluated an alternative method to quantify relaxometry data. Partial least squares (PLS) analysis was applied directly to a variety of simulated time-domain relaxation data under diverse conditions to predict constituent content and results were compared to the standard analysis methods. For many situations, PLS analysis displayed superior performance for quantification than the standard analyses. The technique consistently produced better predictions at lower signal to noise. This robustness to noise makes it an appealing alternative for analysing data from applications that typically have low SNR, such as one-sided sensors, surface measurements, or well-logging. The method also enabled quantification of relaxation rates too close to be separated by an inverse Laplace transform. This capability may allow quantification to be performed using only one-dimensional relaxation data where multi-dimensional measurements were previously necessary to provide constituent separation. The method also enabled quantification of relaxometry data without the need for human interpretation or prior knowledge of what relaxation time is associated with a given constituent. These advantages make PLS analysis an appealing alternative for quantification of relaxometry data in many situations.