Toward Continuous Learning for Real-Time Edge Analytics in Drilling Optimization: A Comparative Study on ROP Prediction

Abstract Regression-based prediction remains a cornerstone of drilling automation, yet its practical deployment is often con- strained by poor generalization and limited adaptability under dynamic field conditions. While many static models achieve high retrospective accuracy, their performance typically degrades in streaming environments, limiting their applicability for edge analytics and downhole monitoring. This study benchmarks representative static and continuous-learning regres- sion models using the standardized Volve USROP dataset. Three model architectures—Support Vector Machine (SVM), Artificial Neural Network optimized with Self-Adaptive Differential Evolution (ANN + SaDE), and Deep Neural Network (DNN)—are reproduced from prior literature and extended with continuous-learning capabilities to emulate real-time op- eration. The evaluation across seven wells examines predictive accuracy, adaptability, and computational latency to assess the feasibility of edge-compatible implementation. Results show that continuous-learning models maintain competitive accuracy while introducing real-time adaptability and low-latency performance suitable for on-site analytics. Among all architectures, the continuous DNN achieved the best overall balance with a Mean Absolute Error (MAE) of 1.80 m/h, coefficient of determination (R2) of 0.87, and an average inference latency of 1.5 × 10−4s per record, demonstrating strong potential for embedded deployment. This work contributes reproducible insights toward developing efficient and trustworthy continuous-learning frameworks for next-generation edge-intelligent drilling automation.