An Unsupervised Learning Algorithm to Compute Fluid Volumes From NMR T 1 -T 2 Logs in Unconventional Reservoirs

A key objective for formation evaluation in unconventional reservoirs is to estimate reservoir quality by quantifying the volumes of different fluid components. Spectroscopy-based tools can estimate the total organic carbon in a reservoir. Resistivity and dielectric tools are sensitive to the water-filled porosity. On the other hand, nuclear magnetic resonance (NMR) tools have the capability and sensitivity to further partition the hydrocarbon and water into fluid components based on their properties and location in the pore space. T-1-T-2 maps from NMR logging tools show unique signatures for hydrocarbons, such as gas, bitumen, and producible and bound oil. Similarly, capillary and clay-bound water and water in larger pores have different signatures. These signatures depend on many factors: properties of the fluids (composition, viscosity), properties of the rock (pore geometry) and the geometrical configuration of fluid phases within the pore space. Unless the fluids have very distinct and nonoverlapping signatures in the T-1-T-2 domain, it is challenging to visually separate the contribution from different fluids and estimate fluid volumes from T-1-T-2 maps. This problem was addressed by an automated unsupervised learning algorithm called blind-source separation (BSS), wherein the NMR T-1-T-2 maps of an entire logged interval are factorized into two matrices: the first matrix contains the T-1-T-2 signatures of the different fluids, and the second contains the corresponding volumes. This method has been shown to work well on multiple field datasets, where there was a sufficient dynamic range in the underlying volume fractions. In this paper, we address two well-known limitations in the BSS algorithm. First, the algorithm assumes a dynamic range in the volume fractions. For this reason, the entire logged interval is considered in the matrix factorization. However, doing so mixes the effects T-1-T-2 maps due to changes in rock properties with changes in fluid volumes. Second, it assumes that the number of sources (or fluids) is known a priori. This is a well-known ill-conditioned problem. We propose several modi cations in the algorithm to address the above limitations. First, we leverage the information that the NMR signature of a fluid is expected to be connected in the T-1-T-2 domain. Second, we assume that each point in T-1-T-2 space corresponds, at most, to one fluid. Lastly, we propose a quantitative metric to guide the analyst in selecting the number of components. We demonstrate the application of this method on simulated datasets as well as field datasets from the Eagle Ford formation and Permian Basin.