Abstract: Pool boiling is an important heat transfer mode in reactor heat exchangers, passive residual heat removal systems, and accident mitigation processes. Accurate prediction of its heat transfer coefficient is essential for evaluating thermal safety margins in nuclear energy systems. However, pool boiling heat transfer is affected by coupled factors such as heating surface inclination, liquid subcooling, and wall superheat. Owing to the limited applicability of their assumptions and calibration databases, conventional empirical correlations often show considerable errors under complex operating conditions, especially when subcooling and inclination effects coexist. To improve prediction accuracy while retaining the physical basis of classical correlations, a residual learning method with an empirical-correlation prior for predicting the pool boiling heat transfer coefficient of water was proposed in this study. A comprehensive experimental database was adopted from 67 independent datasets, including 4 574 data points covering liquid subcooling from 0 K to 85 K, heating surface inclination from 0° to 180°, and heat flux from 1.3 kW/m² to 8 610.5 kW/m². The Han correlation, which considers both subcooling and inclination effects, was selected as the prior baseline model. Machine learning models were trained to learn the residual between experimental data and correlation predictions. Three input-parameter combinations and five representative regression models, including Random Forest, Extra Trees, XGBoost, LightGBM, and multilayer perceptron artificial neural network, were compared. Pure machine learning models were also established to examine the role of the empirical prior. The results show that input-parameter selection significantly affects model performance. Using only heating surface inclination, liquid subcooling, and wall superheat leads to relatively large errors. Introducing heat flux substantially improves prediction accuracy, while further inclusion of the bubble Reynolds number and subcooling Jakob number provides only limited improvement. Among the five models, Extra Trees shows the best overall performance, with a test-set R² of 0.998 5, an adjusted R² of 0.998 5, a mean absolute relative error of 2.83%, and a mean relative error of 0.51%. Compared with pure machine learning models, the empirical-correlation-prior approach further reduces prediction errors and systematic bias, especially for the neural-network model. Error-distribution analysis indicates that the proposed model maintains low errors across different inclination, subcooling, and wall-superheat ranges, with large deviations mainly associated with a few outlier samples. The proposed method combines the physical interpretability of empirical correlations with the nonlinear correction capability of machine learning, providing a robust approach for predicting the pool boiling heat transfer coefficient of water under complex operating conditions.