A decoupling-learning-based unified predictor for dynamic rock strength under sparse and heterogeneous data

Predicting the dynamic strength of rock masses under complex disturbance loads is crucial for safe design of deep underground engineering. However, this task is fundamentally challenged by heterogeneous feature spaces of intact versus fractured rocks, severe data sparsity and measurement noise. To address these issues, this paper proposes a novel decoupling-learning-based unified predictor (DUP). The architecture decomposes the complex mechanical response into two physically transparent sub-tasks: a baseline network capturing intrinsic rate-dependent constitutive behavior, and a correction network that models fracture-induced structural damage. This decoupled learning strategy enables lossless integration of heterogeneous geomechanical data. To overcome sample sparsity, mechanics-informed data augmentation methods are employed to generate physically plausible virtual samples guided by rock mechanics principles. Furthermore, a physics-informed hybrid loss function explicitly embeds domain priors, suppressing noise and enforcing theoretical consistency. Validated on impact test data of synthetic rock-like specimens, DUP consistently outperforms advanced baseline algorithms, achieving superior prediction accuracy. More importantly, the decoupled architecture and physical constraints facilitate interpretability, supporting its reliability for engineering applications.