History matching of a channelized gas reservoir using ES-MDA integrated with an unsupervised serial denoising autoencoder

This study integrates Ensemble Smoother with Multiple Data Assimilation (ES-MDA) with Serial Denoising AutoEncoder (SDAE) for improving the history-matching performance in channelized subsurface environments following non-Gaussian rock facies. SDAE, an unsupervised neural network, is designed as dual autoencoders connected in sequence. A training set composed of channel models is generated using a geostatistical technique called Single Normal Equation SIMulation (SNESim). SDAE is trained by correlating the original channel models as outputs of the neural network and the noise-added models as inputs. The trained SDAE renders channel images that are tuning parameters at each assimilation of ES-MDA. The rendering aims at facilitating the preservation of geological plausibility of the channelized geomodels regarding channel connectivity, width, and pattern. A contribution of SDAE to ES-MDA is tested with application to history matching of a channelized gas reservoir with aquifer support. In this case study, SDAE is coupled with five ES-MDA algorithms, separately: ES-MDA; ES-MDA with Discrete Cosine Transform (DCT); ES-MDA with K-Singular Value Decomposition (K-SVD); ES-MDA with DCT and iterative K-SVD; and ES-MDA with autoencoder. History-matching performances of the coupled algorithms are compared with those of the uncoupled algorithms. Results of the case study highlight the coupling effects that reveal a remarkable recovery of geological plausibility after data assimilation complete provided that sufficient training of SADE is achieved.