Seismic history matching is typically a high-dimensional and severely non-linear inverse problem, which is extremely time-consuming especially for ensemble-based methods. In this work, we propose a deep convolutional auto-encoder (DCAE) to sparsely represent the seismic data and reservoir models using latent features, then perform data assimilation in the low-dimensional model and data space, and finally decode the updated latent features into the original model space. The proposed method provides three main advantages: first, it would less computing time and memory; second, it is helpful to avoid the undesirable phenomenon of ensemble collapse that is a common in high-dimensional data assimilation; thirdly, with the powerful ability in in exploiting spatial and non-linear correlations, DCAE is able to transform the multi-mode data or model space into a lowdimensional space that is approximately normal distributed, which is a basis assumption of data assimilation methods derived from Kalman Filter. The proposed methodology has been tested on a synthetic case of channelized reservoir. It provides a more reasonable history-matched result in terms of reservoir characterization.
Presentation Date: Wednesday, September 18, 2019
Session Start Time: 1:50 PM
Presentation Start Time: 2:15 PM
Presentation Type: Oral