A machine learning method for quantitative characterization of multifluid components based on NMR T1-T2 spectra in shale

Shale reservoirs are complex systems of multifluid components, with abundant oil and gas resources. The low quality of nuclear magnetic resonance (NMR) T1-T2 spectra obtained from shale reservoirs hinders the identification and volume calculation of multifluid components. As a result, the relationship between T1-T2 spectra and multifluid components remains unclear. In this paper, a morphology-based strategy is first proposed for guiding the acquisition of downhole cores, and a training set with rich fluid features can be acquired with a very small number of core samples. Considering the spatial information on T1-T2 spectra, the relationship between T1-T2 spectra and the volumes of multifluid components is established by convolutional neural network (CNN). Based on the prediction results of the trained neural network, Gaussian mixture model (GMM), an intelligent clustering algorithm, is further utilized to identify multifluid components in shale reservoirs. A set of formation with typical shale fluid components is processed to validate the effectiveness of the proposed method. The results show that Image Euclidean Distance outperforms Cosine Distance in indicating the difference between T1-T2 spectra. Based only on a small-sized dataset obtained by morphological processing, CNN can train a neural network with superior predictive performance with the assistance of early stopping. Additionally, GMM effectively reduces the interference signals in T1-T2 features of multifluid components, and the clustering results are more suitable for feature identification of multifluid components in shale.