Long Short-Term Memory and Variational Autoencoder With Convolutional Neural Networks for Generating NMR T2 Distributions

Downhole nuclear magnetic resonance (NMR) logs acquired in the borehole environment are valuable for subsurface characterization because they contain information about the pore size distribution, fluid composition, fluid saturation, fluid mobility, formation permeability, and porosity. NMR log acquisition can be challenging due to operational and financial constraints. Recently, NMR T2 distributions of the subsurface were generated by processing conventional well logs using deep-learning neural-network (NN) models. This improves the accessibility to subsurface pore size distributions. We implement two neural-network models, variational autoencoder-based NN with convolutional layers and long short-term memory (LSTM), to generate NMR T2 distributions from formation mineral content and fluid saturation logs. Prediction performance is evaluated for the entire NMR T2 spectrum ranging from 0.3 to 3000 ms as well as for T2 spectra within four bins obtained by dividing the entire spectrum into four equal parts. Each bin represents a specific pore size range. In terms of R-2, both the models have prediction performances above R-2 of 0.75 for the entire spectrum. The best prediction performance is achieved for the bin of size ranging from 2.7 to 28 ms, representing pore diameters from 10 to 100 nm. The performance of this bin in terms of R-2 is 0.78. The LSTM model is highly sensitive to noise in T2 distribution used during the training, and both the models are robust to noise in the conventional input logs.