PSI - Issue 80

Sadjad Naderi et al. / Procedia Structural Integrity 80 (2026) 77–92 Sadjad Naderi et al. / Structural Integrity Procedia 00 (2025) 000–000

90 14

4.4. Comparative assessment with a purely data-driven model (Gaussian Process Regression) Bayesian updating integrated with a physics-based crack growth model, as implemented in the DBN framework, has been shown in many studies to provide accurate, generalisable predictions while maintaining consistency with established fracture mechanics. The physics-based formulation constrains the learning process, reduces overfitting, and increases interpretability. However, in certain real-world applications where rapid prediction is required, or where structural behaviour is influenced by complex and variable environmental or loading conditions, the added computational effort of physics-based inference can be a drawback. Moreover, when the governing physics are only partially known or deviate from idealised models, these constraints may not fully reflect the true system dynamics. To explore the opposite end of the modelling spectrum, a purely data-driven GPR model is implemented as a benchmark. GPR requires no explicit physics and relies solely on the statistical structure of the data, enabling near-instantaneous prediction. This comparison allows assessment of potential trade-offs between computational efficiency, prediction accuracy, and robustness in interpolation versus extrapolation scenarios.

Fig. 9. Predicted – curves from the GPR benchmark model.