Towards reliable prediction of geopolymer concrete strength across heterogeneous datasets: source-aware validation, uncertainty assessment and whole-mixture calibration

High accuracy is often reported for machine-learning (ML) models of geopolymer concrete, but it is usually estimated on data drawn from the same pool used for model development. This study treats compressive-strength prediction as a source-transfer problem and asks how models trained on heterogeneous public data perform for a material source that was absent during development. Six public datasets were reconciled with their dataset records and parent publications; after screening and exact deduplication, 3,482 observations from 1,597 mixture groups were retained. Seven regression models were evaluated with repeated random, mixture-grouped and similarity-grouped holdouts, nested leave-one-source-out (LOSO) validation, chemistry-proxy ablation, source-aware prediction intervals, applicability-domain and model-disagreement risk analysis, and repeated whole-mixture recalibration. LightGBM achieved a random-holdout mean absolute error (MAE) of 2.78 MPa, which rose to 3.25 MPa when complete mixtures were withheld. Under LOSO validation, the four tree ensembles gave equal-source MAEs of 11.73–12.43 MPa, whereas a global-mean predictor gave 11.27 MPa; for the most shifted source, the zero-shot LightGBM error was 22.68 MPa. With matched hyperparameter tuning, estimated chemistry proxies did not improve macro transfer (11.82 versus 11.88 MPa). Source-aware 90% intervals under-covered the most shifted source, and the nearest-neighbour distance did not rank transfer errors. Recalibration with 5%, 10%, 20% and 40% of the complete target-source mixtures reduced the LightGBM macro MAE from 12.13 MPa to 7.41, 6.07, 4.79 and 3.50 MPa over 20 repeated selections, although the gain varied between sources. Random-split accuracy is therefore an in-distribution (interpolation) estimate where the ensembles did not clearly outperform the global-mean predictor for an unseen source, and the transfer results describe the six studied source datasets. The resulting source-aware workflow, which combines source-held-out evidence, representative local calibration, empirical uncertainty and independent confirmation mixtures, is promising but still requires prospective validation on independent sources.