Machine Learning Classifier for Detecting Non-1D Features Using Deep Directional Resistivity Measurements

Currently, most geosteering jobs can rely on 1D inversion to accurately capture the properties of multiple planarly layered formations in real time. However, when complex non-1D features are present in the formation, higher dimensionality inversions will be required, which need intensive computation resources and more data channels containing 3D information sent to the surface. When abnormal results from 1D inversion are observed, switching telemetry rates and starting higher-order inversions is usually late, which inevitably delays the real-time geosteering decision. To address this, we have developed a machine learning classifier capable of early detection of non-1D features in the formation before the drill bit reaches the higher dimensional feature. A large number of synthetic formations were generated with and without non-1D features that included faults, water coning, and sand injectites, etc., and then fed into a 3D forward model to generate the deep directional resistivity measurements. We trained the classifier to distinguish between 1D and non-1D cases with high accuracy, specifically employing a multilayer perceptron classifier, which is both accurate and suitable for downhole deployment. In practice, this classifier can signal the presence of non-1D features, dynamically select telemetry channels, and initiate specialized inversions in real time. Synthetic and field testing validated its effectiveness, achieving over 80% accuracy when the tool is even 10 m from a non-1D feature, with increasing accuracy as the tool approaches the feature. Field results aligned with geological interpretations, affirming the model's robustness and accuracy. This innovation enhances the ability to recognize complex formations during drilling, automating non-1D inversions and improving bandwidth management in real time, which ultimately increasing drilling speeds. This is one step in the journey towards autonomous geosteering.