Lithology prediction in the subsurface by artificial neural networks on well and 3D seismic data in clastic sediments: a stochastic approach to a deterministic method

A small area covered by a seismic volume was selected for the analysis of using artificial neural networks for the purpose of lithology modelling in a stochastic approach to an otherwise deterministic method. Subsurface lithology was simplified to three categories (sandstone, marl and coal) in accordance with the general geological composition of the Pannonian age sediments in the eastern part of Drava Depression. Two approaches to artificial neural networks were used—training and prediction with a large number of networks with different architecture, and with the same architecture but with the variability of dataset distribution of cases for error calculation in the learning process. Out of a 1000 total cases, 100 realizations of each approach were singled out upon which the data points with probability of 50%, 75% and 90% of occurrence of certain lithology category were upscaled in the model. Six models were generated by indicator kriging. Although in theory, the higher accuracy data should provide a more accurate result, the geologically most sound results were obtained by 50% accuracy data. In higher accuracy results, sandstone lithology was unrealistically over emphasized as a result of the upscaling process, variography and statistical analysis. Presented research can be used in all geoenergy-related subsurface explorations, including hydrocarbon and geothermal explorations, and subsurface characterization for CO 2 storage potential and underground energy storage potential as well.