A physics-boosted transfer learning framework for fracturing pressure prediction with scarce data

Abstract

Accurately predicting fracturing pressure is critical for optimizing the safety and efficiency of hydraulic fracturing operations, particularly in newly developed blocks where data scarcity poses significant challenges. Traditional machine learning methods require large, high-quality datasets to train algorithms. To address these limitations, this study presents physics-boosted transfer learning frameworks designed to enhance fracturing pressure prediction in data-scarce scenarios. By integrating a gated recurrent unit (GRU) deep learning model with physical modeling principles, three transfer learning frameworks were developed and evaluated, including a pure data-driven framework, a hybrid-modelling framework, and a physics-informed framework. Field data from only three shale gas wells were utilized to train the GRU algorithm – simulating real-field data-scarcity scenarios. Fine-tuning technologies are optimized based on the pure data-driven framework. The physics-informed framework demonstrated superior performance, achieving root mean square errors (RMSE) as low as 2–3 MPa, significantly outperforming both the pure data-driven and hybrid frameworks in terms of accuracy, stability, and adaptability. By bridging the gap between data-driven methods and physical modeling, this new framework offers a robust solution, for improving operational safety and cost-effectiveness in hydraulic fracturing, particularly under data-scarce conditions.