Automated Facies Classification for Seismic Inversion

We introduce an algorithm for simultaneous facies classification and fitting of rock physics models from multivariate well log data. Special features of the methodology are designed to render it resilient to data outliers. The algorithm is a robustified and globalized variety of the expectation-maximization algorithm, using reweighted robust nonlinear regression steps for the maximisation step, and heavy-tailed distributional models for the expectation step. Facies classifications are natural byproducts of the expectation step, and optimised rock physics models are produced by the maximisation step. The practical advantages of the approach are illustrated using data from the Satyr-5 well, located in the Northern Carnarvon Basin, North West Shelf of Australia. Outputs of the algorithm include facies labels and free parameters in the corresponding rock-physics models, which can be easily interpreted and directly used in downstream workflows such as facies-based seismic inversion.