NMR Logging Data Processing Using Deep Learning

Abstract

Nuclear magnetic resonance (NMR) is a powerful tool in biomedical, chemical analysis, the oil industry, and other scientific fields. It provides information on molecular structure for the analysis of molecular dynamics and interactions. In recent years, deep learning (DL) has attracted great interest in various research fields because of the availability of high-performance computing. The employment of DL methods to effectively address shortcomings in NMR data processing is a new research field, such as signal reconstruction, MRI reconstruction, and peak picking of protein spectra. In general, the application of DL in NMR can be further summarized in three aspects: high signal-to-noise ratio (SNR) NMR signal reconstruction, high-resolution spectral reconstruction, and automatic interpretation of spectra. Inspired by these successful applications, we consider that DL can be applied to the construction of low-field NMR relaxation spectra, which would help to accurately characterize the structure and properties of rock porous medium. Low-field NMR techniques use Carr-Purcell-Meiboom-Gill (CPMG) pulse sequences to accurately measure the formation. The measured echo signals always have strong noise and are used to determine the T2 spectra by inverse Laplace transform (ILT). T2 spectra provide information on pore structure, fluid saturation, and permeability to further evaluate reservoirs. However, the ILT process is ill-conditioned, and the solutions are not unique. The results may reduce the resolution of spectra, thus influencing subsequent interpretation and application. To overcome these problems, we proposed the signal denoising framework and spectra inversion network based on DL. First, an NMR forward simulation process was implemented to generate data sets for DL model training. Signal parameters and Gaussian distribution are regarded as prior knowledge integrated into the simulation process. Second, the autoencoder network was trained by a forward simulation data set to remove the noise contained in the signals. After compression and reconstruction, signals with high SNR can be obtained. Finally, we design the attention multiscale convolutional neural network (ATT-CNN) for spectra inversion. The energy changes of the signals were extracted by the attention mechanism. A multiscale convolution neural network (CNN) is designed to extract the local noise fluctuations and global attenuation characteristics of the echo signals. Simulated data and rock core data measured in the laboratory were used to verify the effect of DL models. The NMR signals were input into the autoencoder model for denoising. The output signals with high SNR were going through the ATT-CNN model to obtain the T2 spectra. The traditional inversion method based on regularization is also used as a comparison. The result demonstrated that the ATT-CNN model could be more adapted to low SNR echo signals and inverse more sparse and stable spectra. Meanwhile, the pore-size distribution and porosity of rock cores could also be accurately characterized based on high-resolution spectra. This work shows that prior knowledge constrained to the data set and network model can make inverse spectra more accurate. DL methods can become a powerful tool for revealing the structure and properties of the porous medium. We hope that this optimization scheme inspires more applications, especially in formation evaluation.