Seismic reservoir characterization of the Gassum Formation in the Stenlille aquifer gas storage, Denmark - Part 2: Unsupervised classification

Ideally, a good static reservoir model provides an accurate estimate of the extent, porosity, permeability, and lithology of the container as well as the properties of the seal and any faults or fractures that may allow the reservoir to leak. The major pitfall of deterministic, statistical, or supervised learning workflows is that they estimate only the properties sampled by the wells or provided by empirical relations and may miss mapping heterogeneities in the reservoir and seal that can give rise to flow baffles and reservoir leakage, in contrast to supervised. This shortcoming is exacerbated when the number of wells is small, the types of logs recorded limited, and the migrated seismic gathers absent or of limited quality. In contrast, unsupervised learning looks for patterns in the seismic amplitude and attribute volumes themselves. In this paper, we apply unsupervised learning algorithms to evaluate the natural gas storage Stenlille aquifer in Denmark and compare the results to a supervised multiattribute regression reservoir characterization described in a companion paper. Specifically, we apply principal component analysis, self-organizing mapping, and generative topographic mapping workflows to extract patterns across eight attribute volumes: relative acoustic impedance, energy, sweetness, GLCM entropy, curvedness, and three spectral magnitude volumes. We find that the large-scale patterns are similar, but that the unsupervised learning algorithms provide greater detail. Because our deterministic model was built on poststack data using the limited well log data available, we believe that the heterogeneity mapped by the unsupervised learning workflows provide a relatively unbiased means of estimating risk in our reservoir model. Quantifying the importance of these anomalies will need to be reconciled with a dynamic reservoir model.