Robust collapse prediction for corroded subsea pipelines under multiaxial loading: A physics-informed Bayesian framework constrained by symbolic regression

Subsea pipeline collapse under multiaxial loads and corrosion severely threatens offshore structural integrity. This study introduces an integrated FE-SR-PINN-BNN framework, combining detailed 3D finite element (FE) numerical modeling in ABAQUS with a physics-informed Bayesian neural network (PINN-BNN) constrained by symbolic regression (SR). Explicit algebraic expressions of underlying mechanical mechanisms derived via SR are embedded into the PINN-BNN loss function as structural physics constraints. Numerical results show that the proposed model achieves a high coefficient of determination (R 2 = 0.9940) with prediction residuals within ±1.5 MPa, significantly outperforming purely data-driven machine learning models and empirical formula-constrained physics-informed baselines. In robustness evaluations with 5%-20% added noise, its R 2 performance degrades by only 33.49%, substantially lower than benchmark model drops. Under an extreme small-sample evaluation using only 10% training data, its R 2 reaches 0.8464, outperforming four purely data-driven baselines by 97.99%, 42.18%, 39.44%, and 8.35%, respectively. Beyond precise collapse pressure predictions, the framework incorporates uncertainty quantification based on 95% prediction intervals, achieving a mean prediction interval width of 1.88 MPa and a coverage probability of 0.952, exceptionally close to the theoretical 0.95 target. This study provides reliable support for structural integrity and safety assessment of subsea pipelines in complex service environments.