Novel application of artificial neural networks to derive lithofacies in the Bunter Sandstone Formation of the UK Southern North Sea

The Bunter Sandstone Formation (BSF) in the UK sector of the Southern North Sea is thought to have a significant potential for CO₂ storage, which would help the UK achieve net-zero carbon emissions by 2050. During the assessment phase, a robust lithofacies classification scheme and accurate identification enable better control for delineating the petrophysical property distribution in 3-D space, which is vital for further estimating the storage capacity and simulating CO₂ migration.   

In previous studies, several different lithofacies classification schemes were proposed for calculating the CO₂ storage capacity of BSF traps. However, the establishment of these schemes was almost entirely dependent on well-logging data due to limited cores and corresponding thin sections. For example, the ‘cemented sandstone layers’ were identified only by low gamma-ray values with a sharp increase in density values and a decrease in acoustic values, which leaves significant uncertainty because of the lack of detailed lithological description and core calibration. For lithofacies identification based on well-logs, artificial neural networks are of great potential due to their strong non-linear mapping ability. Numerous researchers used the fully connected neural network (FCNN) to recognize lithofacies, but this method can only construct point-to-point mapping, which cannot take into account the previous information (data points in well-logs) of sequence data and results in not being fully competent in lithofacies identification.

This study aims to partition the BSF reservoirs into several relatively homogeneous lithofacies based on cores, thin sections, SEM (Scanning Electron Microscope) and XRD (X-ray Diffraction) analysis. We summarize each lithofacies characterization including grain size, porosity/permeability, 3D network structure of pore and cement determined by X-ray CT. Data pairs composed of logs and corresponding lithofacies types are selected for training neural networks. On this basis, we employ the algorithms designed for sequence data to achieve lithofacies identification, and make a comparison with the widely used FCNN method.