Classification of Drilled Lithology in Real-Time Using Deep Learning with Online Calibration

Abstract

Decision making to optimize the drilling operation is based on a variety of factors, among them real-time interpretation of drilled lithology. Because logging while drilling (LWD) tools are placed some meters above the bit, mechanical drilling parameters are the earliest indicators, although they are difficult to interpret accurately. This paper presents a novel deep learning methodology using mechanical drilling parameters for lithology classification. A cascade of deep neural networks (DNNs) are trained on historical data from wells on a field operated by Equinor. Rather than an end-to-end approach, the drilling parameters are used to estimate LWD sensor readings in an intermediate step using the first DNNs. This allows continuous updates of the models during operation using delayed LWD data. The second DNN takes the virtual LWD estimates as input to predict currently drilled lithology, similar to manual expert interpretation of logs. This configuration takes into account case-dependent [mud, bottomhole assembly (BHA), wellbore geometry] and time-varying (bit wear, wellbore friction) relationships between drilling parameters and LWD readings while assuming a constant rule when using LWD data to classify lithology. Upon completion of training and validation, the system is tested on a separate, unseen wellbore, for which results are presented. Visualizations for true lithology alongside the estimates are given, along with confusion matrices and model accuracy. The system achieves high accuracy on the test set and presents low confusion between classes, meaning that it distinguishes well between the lithologies present in the wellbore. It can be seen that the borders between successive layers of lithology are detected rapidly, which is crucial seen from an optimization standpoint, so the driller may immediately adjust accordingly. It shows promise as an advisory system, capable of accurately classifying currently drilled lithology by continuously adapting to changing downhole conditions. Although we cannot expect perfect estimates of lithology purely based on drilling parameters, we can obtain a preliminary map of the subsurface this way. This novel configuration gives a real-time interpretation of the currently drilled lithology. Thus, the drilling operation can be improved through early information and prompt drilling parameter adjustments to accommodate the current drilling environment.