Tying detailed well log measurements to lower resolution but a really extensive 3D seismic data volumes is key to quantitative seismic interpretation. Ties using a poststack or prestack convolution model are routine, while supervised classification tying well data to seismic attributes using neural networks and geostatistics are also well established. However, unsupervised classification ties where the objective is to identify unknown patterns in the data is less well established. In this paper, we use an automatic learning Gaussian Mixture Model to statistically characterize the well logs, evaluate the probability distribution functions of different lithologies and then tie them to corresponding 3D seismic attribute volumes. We precondition our four-dimensional data by projecting onto two dimensions using Independent Component Analysis.
We apply this workflow to Diamond M Field within the Horseshoe Atoll Reef Complex, Scurry County, TX, and find the Gaussian Mixture Model is able to statistically characterize and resolve lithological variations seen in the logs. In particular, we are able to clearly distinguish between lithologies from six different wells in the region of interest. The final result is a probabilistic map that statistically measures the variability of the seismic lithologies from the well logs.