Physics-Informed Multitask Learning with Dynamic Weighting for Efficient Prediction of Extended-Reach Wellbore Cleanliness

Summary Inadequate hole cleaning can lead to downhole complications such as stuck pipe, making the prediction of cuttings transport behavior in extended-reach wells critically important for drilling operations. However, traditional experimental and numerical simulation methods face limitations in adaptability and computational efficiency, while purely data-driven models often suffer from poor generalization capability. To address these challenges, we propose a dynamically weighted physics-informed multitask learning (PIMTL) model. Based on the two-layer cuttings transport mechanism, the governing equations are incorporated as physical constraints into the data-driven model, enabling efficient and accurate prediction of multiple cuttings distribution characteristics in extended-reach wells. The multitask learning (MTL) model is constructed to predict multiple strong coupling parameters related to cuttings transport, including the height of the cuttings bed, pressure drop per unit length, velocity of the suspension layer, and velocity of the moving cuttings bed. The PIMTL model integrates gradient normalization (GradNorm) and a physics-constrained annealing strategy to dynamically adjust task weights. The model’s predictive accuracy and generalization capability are evaluated under various working conditions and across different data sets. Ablation studies are conducted to assess the contributions of individual components within the PIMTL framework. The results indicate that, compared with single-task models, the collaborative prediction capability of MTL not only improves prediction accuracy but also significantly enhances computational efficiency, with the average coefficient of determination (R2) increasing by 6.15%, the average mean absolute percentage error (MAPE) decreasing by 53.48%, the number of parameters reducing by 22.18%, and the computation time reducing by 44.62%. After applying physical constraints, compared with purely data-driven models, the PIMTL model achieves superior prediction accuracy. For the prediction of the height of cuttings bed, the R2 value reaches 0.9973 and the MAPE is as low as 0.16%. It is also applicable to cases involving smaller-scale data sets. Moreover, the model demonstrates better generalization performance under varying operational conditions and across different data sets. This study provides effective technical support for the accurate and efficient evaluation of hole cleaning in practical drilling engineering and serves as a foundational framework for real-time decision-making systems in drilling operations.