Mechanism-Aware π Group-Based Gene Expression Programming for Axial Capacity Prediction of Steel-Concrete Composite...

Reliable engineering prediction requires accuracy, interpretability, physical consistency, and uncertainty awareness. Concrete-filled circular steel tubes (CCFSTs) axial capacity is a representative case because it is governed by coupled material, geometric, confinement, and instability effects. Existing formulas are transparent but often inflexible, whereas black-box machine learning models can be accurate but are difficult to audit directly. This study proposes a physics-guided symbolic model (PI-ResGEP) that combines dimensionless π-group regression, mechanism-aware feature selection, residual diagnosis, and conformal uncertainty quantification. The explicit backbone is the primary predictor, while the residual layer is framed as a conservative correction and diagnostic component. A primary dataset comprising 833 CCFST tests was compiled from 55 published studies, while 313 data points collected from 34 additional studies were used for external validation. Experimental results show competitive accuracy with an R² of 0.9946 while retaining a directly inspectable equation. By integrating distribution-free conformal prediction, the model achieves 90.8% coverage at the 90% nominal level and identifies heteroscedastic uncertainty regimes. PI-ResGEP may provide a preliminary predictive reference for cases beyond the applicability limits of Eurocode 4 and offers a potentially transferable workflow for engineering prediction problems requiring physically consistent, interpretable, and uncertainty-aware estimates.