In the petroleum industry, the characterization of reservoir properties is essential for designing optimal development plans. Although geostatistical modeling is effective for estimating distributions of reservoir properties, it becomes difficult to estimate them accurately if the well data (hard data) is sparsely distributed. To solve this problem, we focused on deep learning.
The objective of this study is to investigate the usefulness of soft data created by deep learning in reservoir characterization. To accomplish this objective, we developed two programs. Program-1 was constructed using a Convolutional Neural Network (CNN). It has a function to learn well data (permeability data at wells) that have been prepared for four groups of different sedimentary environments in advance, and to classify the well data taken from a target field into the appropriate sedimentary environment group. Program-2 adopts the template matching and selects the permeability distribution, which is the most similar to the well data, from the dataset stored in the group specified by Program- 1. The permeability distribution data thus selected by Program-2 is then used as soft data. We tried to estimate the permeability distribution of a hypothetical reservoir. The permeability distributions were estimated by Sequential Gaussian Simulation (SGS) only with the well data, and by Kriging with External Drift (KED) and Co- Kriging (CK) using the soft data selected by Programs-1 and -2. In this study, the permeability distributions estimated by KED and CK showed better agreements with the true data compared to those estimated by SGS.
The above method can be successfully applied for the dataset including the soft data of permeability distribution similar to the true one. Therefore, we are now modifying Program-2 to draw the multiple images of facies distribution constrained by well data using Generative Adversarial Network (GAN).
