Enhanced self-adaptive physics-informed neural networks for prediction of wellbore fluid thermal dynamics during deepwater drilling

Real-time and accurate prediction of transient downhole temperature is critical for safe and efficient deepwater drilling. However, conventional numerical simulators are computationally expensive, and field measurements are often sparse and noisy. This study introduces an enhanced self-adaptive physics-informed neural network framework (ESA-PINNs) for wellbore transient heat transfer modeling, in which governing equations, boundary conditions, and observations are integrated through a multi-source loss function. To improve training efficiency and stability, the framework incorporates residual-based adaptive sampling and self-adaptive loss weighting, and further adopts a fine-tuning strategy to reuse pretrained models and accelerate convergence under changes in operating and thermal conditions. Numerical experiments demonstrate that ESA-PINNs outperform conventional PINNs and data-driven baselines in convergence speed, robustness, and error reduction, achieving reliable predictions under highly sparse observations with mean absolute percentage errors below 1%. Moreover, k-fold cross-validation (k = 5) is employed to benchmark ESA-PINNs against six machine learning algorithms, and a field case study validates the practicality of the method, showing its capability to assimilate noisy and irregular downhole data while preserving physical consistency and reproducing temperature distributions consistent with measurements. Future work will extend the framework toward parameterized PINNs to enable well-to-well prediction under larger variations in geometry and drilling configurations.