Azimuthal LWD Data Interpretation for UBCTD Geosteering Using a Physics-Informed Neural Network

Abstract

The economic success of horizontal wells in complex carbonate reservoirs is profoundly sensitive to precise wellbore placement within narrow target zones. Conventional geosteering, which relies on the real-time subjective interpretation of Logging-While-Drilling (LWD) data, is susceptible to human bias and often leads to suboptimal decisions. While data-driven machine learning offers an alternative, purely statistical models frequently produce physically implausible predictions. This paper introduces a novel Physics-Informed Neural Network (PINN) framework that seamlessly integrates domain knowledge with a deep learning architecture to automate and enhance geosteering classification. The methodology employs a multi-layer perceptron trained with a custom composite loss function, which augments standard cross-entropy loss with two physics-derived penalty terms: one enforcing petrophysical consistency between sensor readings and predicted steering actions, and another promoting wellbore trajectory smoothness. Trained and validated on a sophisticated synthetic dataset engineered to replicate the geological complexities of a Central Asia Shu formation carbonate reservoir, the model demonstrates a significant performance improvement over a purely data-driven baseline. The refined PINN model achieved a test accuracy of 91.18%, a substantial increase over the baseline Logistic Regression model's 86.68%. Crucially, the physics-informed constraints led to a dramatic enhancement in recognizing the optimal 'stay' condition, with the F1-score for this critical class rising from 0.01 to 0.51. An ablation study confirmed the petrophysical constraint as the primary driver of this improvement. Post-hoc explainability analysis using LIME (Local Interpretable Model-agnostic Explanations) verified that the model's decision-making aligns with established geological principles. This work successfully demonstrates that embedding physical constraints directly into the ML learning process yields a more robust, reliable, and interpretable autonomous geosteering system, paving the way for more objective, data-optimized well placement and reduced reliance on subjective interpretation.