Inverse Modeling with Deep Neural Network and K-medoids Clustering Under Uncertain Geological Scenarios

This paper presents an inverse modeling based on deep-neural-network, of which scheme integrates data-encoding with stacked autoencoder and k-medoids clustering to select the adequate geo-models for the supervised-training dataset in the presence of uncertain geological scenarios. The reliable geological scenarios are essential at the successful history matching as well as the accurate forecasting but the limited data obstruct the consideration of well-production-performances. The developed method reduces the errors of matching and forecasting profiles as workflow stages, and results out the reliable plausible geo-models satisfying different well-oil-rates. K-medoids clustering screens error-prone geo-models implementing flow-response-based distances. The results show that deep-neural-network can be applicable as a robust history-matching tool under multiple geological interpretations.