Predicting bond-slip behaviour in grouted bellows connect rebar using deep learning

Prefabricated construction has emerged as a pivotal driver of global architectural industrialization and sustainable development, owing to its advantages in efficiency, environmental sustainability, and quality controllability. By integrating factory-manufactured components with on-site assembly, this approach significantly enhances construction productivity while minimizing environmental pollution, aligning with national energy-saving and emission-reduction policies. Notwithstanding these advantages, the structural safety of prefabricated concrete systems critically hinges on joint connection technologies. Among these, grouted corrugated duct (GCD) connections have garnered attention for their construction convenience and reliability. While GCD technology achieves component anchorage through grout-filled ducts, its mechanical performance is directly governed by rebar anchorage effectiveness, necessitating in-depth exploration of parameter impacts on bond stress-slip relationships. Traditional methodologies—such as experimental testing, finite element simulations, and statistical formulations—face limitations in cost, generalizability, and resolution of high-dimensional nonlinear interactions. The advent of machine learning (ML) offers a paradigm shift, enabling data-driven analysis of complex anchorage mechanisms and paving the way for optimized joint design.

In parallel, ML applications in civil engineering have revolutionized GCD connection performance research. Integrated learning models (e.g., XGBoost, Random Forest) demonstrate superior accuracy over empirical formulas in predicting bond stress-slip curves. For instance, Li et al. (Year) trained models on 1,056 experimental datasets to accurately predict mechanical behavior at rebar-grout interfaces, while Pishro et al. (Year) enhanced predictive efficiency by integrating physics-informed neural networks (PINNs) to optimize bond constitutive models for ultra-high-performance concrete. Furthermore, GCD technology has been validated through real-world applications in bridges and shear walls, exhibiting seismic performance comparable or superior to cast-in-place systems. Looking ahead, future research must prioritize ML-driven investigations into multifactorial coupling effects and the development of generalizable constitutive models, ultimately advancing intelligent and standardized design frameworks for prefabricated structural nodes.

This study proposes a CNN-LSTM coupled deep learning model (Figure 1), which integrates temporal feature extraction and long-term dependency analysis capabilities to predict the bond stress-slip constitutive relationship of grouted corrugated ducts (GCDs). The model inputs critical parameters—including grout tensile strength, rebar ultimate strength, and slip displacement—and outputs corresponding bond stress, enabling full-cycle curve prediction. Leveraging a dataset of 114 specimens (2,272 samples) and validated through 15 independent tests, the CNN-LSTM model demonstrates reduced prediction errors compared to empirical models, with significantly enhanced capability to capture complex nonlinear interactions. Four evaluation metrics—total energy dissipation, peak bond stress, failure modulus, and residual energy—were employed to quantify performance.

Comparative analysis using Liu et al.’s experimental data (Figure 2) reveals that the CNN-LSTM-predicted bond-slip curves align closely with test results, whereas empirical formulas exhibit notable deviations in predicting ultimate stress and slip displacement. Further validation via 15 independent tests confirms the model’s superior generalization, particularly under grout strengths ≥5.165 MPa, where its yield stage prediction accuracy surpasses empirical approaches. Sensitivity analysis based on Mean Impact Value (MIV) identifies rebar strength and grout tensile strength as dominant influencing factors, while the corrugated duct aperture ratio exhibits negligible impact.

To address the interpretability limitations of deep learning, a simplified hyperbolic constitutive model is proposed. This model, parameterized using CNN-LSTM-predicted peak points, achieves >90% consistency with experimental ascending curves, while its conservatively designed descending segment ensures engineering safety. Embedded into ABAQUS spring elements for nonlinear simulation (Figure 3), the simplified model demonstrates consistent failure modes with experiments, with total energy dissipation errors <8%, providing a reliable tool for rapid engineering assessment.

This paper ” Predicting bond-slip behaviour in grouted bellows connect rebar using deep learning” was published in Smart Construction.

Yihu C, Xingshuo Y, Jingchao L, Guangxin X, Min Z, Yanwei W. Predicting bond-slip behaviour in grouted bellows connect rebar using deep learning. Smart Constr. 2025(1):0006, https://doi.org/10.55092/sc20250006.

END