Machine learning-assisted conformal uncertainty quantification of punching shear resistance in RC slab–column joints

Punching shear failure in reinforced concrete slab–column joints is safety–critical because it can occur suddenly with limited deformation warning before loss of load-carrying capacity. Probabilistic machine-learning studies have advanced punching shear assessment by modeling resistance uncertainty and failure-mode tendencies. However, for pre-test assessment, it remains necessary to examine whether prediction intervals constructed only from design-available variables provide reliable lower-bound information for brittle punching cases. This study develops an internally evaluated log-conformal interval audit for predicting punching shear resistance in RC slab–column joints. A complete-case database of 594 specimens was used with six inputs: reinforcement ratio, effective depth, equivalent column width, longitudinal reinforcement strength, shear-span/depth ratio, and concrete compressive strength. We compared six backbones: CatBoost, XGBoost, LightGBM, Random Forest, Gradient Boosting, and ExtraTrees. The target resistance was modelled in logarithmic space, and intervals were generated using cross-conformal calibration to obtain positive, asymmetric resistance bounds after back-transformation to kN. The brittle or ductile failure label was not used as a regression input or for interval grouping; it was used for split stratification and post-hoc interval auditing. Across 30 repeated random splits, CatBoost achieved a mean R 2 of 0.9184 ± 0.0770, while the main split yielded R 2 = 0.9724 and was retained for analysis. Repeated-split evaluation indicated that the 90% CatBoost + D-LCI yielded a depth-conditioned general interval, whereas the 95% CatBoost + PR-LCI provided a wider, more conservative lower-bound interval with a lower tendency toward brittle unsafe-miss. The workflow supports lower-bound uncertainty screening within the database and evaluation protocol.