Enhanced AI-Driven Automatic Dip Picking in Horizontal Wells Through Deep Learning, Clustering, and Interpolation, in Real Time

The analysis of borehole image logs is important for subsurface studies but becomes especially crucial when extracting real-time structural information for geosteering in horizontal wells. Indeed, these images help extract data about bedding surfaces, fractures, and faults, which enable the construction of three-dimensional (3D) reservoir models and optimal well placement for future production optimization. Borehole images in horizontal wells are challenging for dip picking—we observe mainly lengthy parallel and ovoid bedding dip traces called “bull’s-eyes,” as the well trajectory may be subparallel to the bedding. This deviates considerably from the classic model of dip picking, which extracts only sinusoids. So far, the delineation of non-sinusoidal bedding features has relied on marking the trace by a series of manually picked segments. In this paper, we present a method that enables the precise automatic extraction of segments from non-sinusoidal features using an artificial intelligence (AI) model and propose an automated grouping mechanism of the segments. Such a solution is applicable in real-time scenarios, facilitating geosteering guidance. Our solution is an automated workflow that detects and picks non-sinusoidal bedding dip traces in real time in horizontal well borehole images and computes the corresponding orientation of the structure. The workflow starts with borehole images and the associated segments provided by the “auto dip-picking” algorithm. A convolutional neural network (CNN) detects bedding features and categorizes them as sinusoidal or non-sinusoidal bedding features. Subsequently, segments are regrouped within each bedding feature, creating comprehensive data sets for each feature. Single-segment sinusoidal features are preserved, while multi-segment ones undergo an advanced clustering mechanism based on orientation and on a derivative of the sinusoidal function associated with the segment. Meanwhile, parallel and bull’s-eye structures undergo a transformative process—a recursive approach connects segments within the same layer. Then, we compute each layer’s global orientation. Our study yielded significant outcomes by automatically detecting non-sinusoidal bedding features and computing associated dips from borehole images in horizontal wells. The integration of our advanced workflow reduced manual intervention. In addition, this workflow is versatile, catering not only to horizontal wells but also to vertical ones. We provide a solution capable of handling simultaneously non-sinusoidal bedding and sinusoidal bedding features automatically with just one click. By embracing automation, we also eliminate subjective interpretations, ensuring a standardized and efficient analysis process.