The rate of penetration (ROP) is a critical statistic for evaluating drilling efficiency, and its accurate prediction is essential for improving drilling performance and reducing operational costs. Conventional physical ROP models are theoretically sound but exhibit limited predictive accuracy, while machine learning algorithms can achieve higher precision yet often lack sufficient interpretability and generalization capability. Therefore, this study presents a physics-informed neural network (PINN)-based hybrid model for ROP prediction and transfer generalization. The proposed approach integrates the mechanical specific energy and Soares ROP models as physical constraints and establishes a hybrid deep learning architecture that combines a squeeze-and-excitation residual network (SE-ResNet), bidirectional long short-term memory (BiLSTM), and self-attention mechanism (SAM). Within this framework, an adaptive loss weighting strategy is introduced into a unified loss function to dynamically balance the contributions of data fitting and physical constraints during training. The experimental results show that compared with traditional physical models and mainstream intelligent algorithms such as support vector regression (SVR), back propagation neural network (BP), LSTM, the PINN model demonstrates superior predictive performance, with a root mean square error (RMSE) of 6.214, mean absolute error (MAE) of 3.483, mean absolute percentage error (MAPE) of 6.102%, and coefficient of determination (R2) of 0.953. Furthermore, to evaluate the modelʼs cross-domain generalization capability, zero-shot and few-shot transfer learning approaches are investigated. The few-shot transfer method achieves a MAPE of 10.413%, confirming the modelʼs adaptability and robustness across different drilling datasets. This study helps to advance the deep integration of artificial intelligence, big data, and other technologies with traditional drilling mechanism models, and has significant theoretical and engineering application value.