Abstract Collapse pressure is an important part of calculating the lower limit of the safe mud density window, which is crucial for optimizing the well trajectory, designing the drilling fluid density, and ensuring drilling safety. However, collapse pressure prediction methods based on physical mechanisms are theoretically complex, computationally intensive, and slow. Data mining-based prediction methods often rely on conventional machine learning models, which suffer from low prediction accuracy, high data demand, and poor interpretability. In this paper, a novel hybrid-driven model combining mechanistic knowledge and machine learning methods is proposed, which has a faster computational speed in collapse pressure prediction compared with the traditional analytical model, and a better performance compared with the existing data-driven models. The model incorporates the stress transformation, rock strength criteria, and other knowledge to ensure the robustness and interpretability of this prediction model. The neural network structure and model hyperparameters are optimized using a Bayesian optimization algorithm. To consider the influence of uncertainty of input parameters on collapse pressure, the Monte Carlo method is used to quantify the influence of uncertainty of input parameters on collapse pressure prediction results based on the hybrid-driven prediction model, and the sensitivity of different input parameters to the outcomes is determined. The proposed model, tested on a test dataset, demonstrated high prediction accuracy and prediction stability with an average absolute error of 0.0037 g/cm3 and a root mean square error of 0.0104 g/cm3 for the collapse pressure equivalent density. Furthermore, a horizontal well was selected for validation, with predicted results exhibiting an average absolute error of only 0.0045 g/cm3 compared to the logging interpretation results, and a computational speed nearly 100 times faster than traditional analytical models. Three points with different stress conditions were selected on this well and their equivalent density of collapse pressure hemispheric projection maps were predicted, and the results were consistent with the actual results, indicating that the model can accurately capture the variation of collapse pressure with well inclination and azimuth. To quantify the effect of input parameter uncertainty on wellbore stability, the influence of input parameter uncertainty on the equivalent density of collapse pressure is simulated based on the above prediction model in combination with the Monte Carlo method, and the corresponding confidence intervals are given. The results found that the effect of uncertainty in ground stress on collapse pressure is relatively significant. In conclusion, the hybrid-driven model effectively integrates physical knowledge, enabling rapid and accurate prediction of collapse pressure in horizontal and inclined wells, offering an innovative approach for intelligent wellbore stability assessment and uncertainty analysis.
6
