The past few years have seen increasing interest in the application of machine learning techniques in the industry, specifically in seismic interpretation. In this work, we benchmarked advanced neural network algorithms against standard probabilistic lithology classifications from seismic data, to understand their benefits and limitations, and to check which approach works best under which circumstance. Based on a well-derived rock physics model in a clastic setting, we tested the ability to predict lithotypes from inverted seismic data using a Bayesian lithoseismic classification, classification using a Democratic Neural Network Association, and the direct neural network inversion for rock properties (PHIE, Vshale).

The presented workflow starts with a conceptual geological model, seismic and well data preconditioning, and the creation of a rock physics model. Then, the prestack seismic data are simultaneously inverted to elastic rock properties, and various seismic pre- and poststack attributes are extracted. Combinations of these attributes are the input into the classification and prediction algorithms and the results are compared and discussed.

The findings of this and other projects show that lithoseismic classification works well in data sets with sparse well information, but in cases of significant overlap of properties, such as carbonates, it has limitations. The DNNA classification requires significantly more well data input for training. It reveals more details even for overlapping facies types, as more training attributes can be used. When sufficient well data is present, the direct inversion for rock properties is an elegant solution for predicting and mapping these rock properties, even if there is a nonlinear dependency on the elastic attributes.