A Cyclic Constitutive Model for Oil-Well Cement Slurries Based on Hybrid Deep Learning Architectures and Physics-Informed Constraints

During injection and production for oil and gas geological storage, the cement sheath is frequently subjected to high-magnitude, high-frequency cyclic compressive loads. These cyclic loads can induce progressive, irreversible damage in the cement sheath in the form of brittle microcracking and plastic deformation, while the casing and surrounding formation typically exhibit predominantly elastic recovery during the unloading phase. However, existing fatigue life prediction models often fail to capture the dynamic stress-strain constitutive behavior of cement sheath under cyclic loading. In this study, intelligent inversion methods (i.e. Artificial Neural Networks) were employed to directly capture highly nonlinear and complex correlations among variables from measured experimental data, offering greater flexibility and adaptability in deriving material constitutive models. Five constitutive prediction models for the complex hysteresis loops of hardened oil-well cement slurries under cyclic loading are developed using deep learning (DL) statistical theory and physics-informed constraint methods. Firstly, to effectively describe the nonlinear morphological evolution of the hysteresis loops in cyclic curves, the experimental dataset is subdivided into 416 sub-datasets according to different cycle periods. Three different hybrid DL architectures—LSTM, CNN-LSTM, and TCN-LSTM—are constructed, and their learning accuracy and effectiveness are evaluated. Then, combined with physics-informed (physics-constrained) supervised approaches, ablation studies are conducted to compare and assess the use of a single-step prediction model with physics-based constraints. Finally, the optimal model is extended to the full-process prediction of the entire cyclic loading paths. The research approach presented in this paper differs from previous methods that input the entire cyclic curve data into an ANN model all at once. Instead, it establishes a novel methodology characterized by single-step rolling learning, enhanced accuracy through physics-informed constraints, and continuous full-process prediction, demonstrating excellent predictive performance (R² > 0.98).