Cross-database validation and calibrated uncertainty for machine learning models of the shear capacity of FRP-strengthened RC beams

Machine learning models for the shear contribution of externally bonded fibre-reinforced polymer (FRP) to reinforced concrete beams are validated almost exclusively on a random hold-out split of a single self-compiled database. Because such databases are assembled from a few dozen experimental programmes, each contributing several near-identical specimens, a random split leaves siblings of nearly every test specimen in the training set. This study provides a cross-database assessment of that practice. Two independently compiled databases (442 and 455 specimens) were harmonised into a provenance-traced dataset of 828 records from 167 programmes spanning 1992–2024, and six learners were evaluated under five regimes of increasing severity. For gradient boosting the coefficient of determination fell from 0.67 and 0.69 under random splitting to 0.39 and 0.22 under study-grouped validation, and to 0.21 under cross-database transfer. A control in which the same physical specimens were scored under each compilation in turn changed little (0.35 versus 0.37), showing that clustering by programme, not curation disagreement, is responsible. A hierarchical errors-invariables model attributes 38% of the residual variance to programme identity and delivers calibrated predictive intervals (91.8% coverage at a nominal 90%), which span a factor of about twelve. Design provisions remain less accurate than the fitted models, but the reported margin roughly halves under honest validation. A reporting protocol is proposed, requiring publication of the specimen-to-study mapping, validation grouped on it, disclosure of out-of-range evaluation cases, and reporting of interval coverage and a characteristic value.