Work

Physics-guided machine learning for leakage-aware prediction and temporal extrapolation of concrete carbonation

Year: 2026
Type: article
Abstract: Machine learning models for carbonation depth are usually validated on random splits of multi-source compilations and seldom asked to predict beyond their training ages. We re-evaluated 17 model configurations on 8,120 natural carbonation depths from 71 sources (7 days to 20 years) under five protocols, separating forecasts for series already observed early from forecasts for new sources. Within-series time exponents were 0.52 indoors, 0.46 sheltered and 0.32 unsheltered; only the unsheltered exponent stayed below 0.5 under source resampling, alternative weightings and a censored likelihood, and the outdoor exponents rose at later ages. Random-row testing overstated accuracy: R 2 of about 0.93 fell to 0.63 to 0.67 on unseen sources, where a physics-normalised learner gained little, and only in a few sources. Training gradient boosting on the coefficient d/ \({\text{t}}^{{{\text{n}}_{{\text{e}}} }}\) restored extrapolation in time: R 2 = 0.87 against 0.17 for later ages of observed series, and 0.57 against at most 0.03 for new sources at late ages. Nested hyperparameter tuning did not change these results. Source-grouped residual bands covered 91.8% of new-source depths, unevenly across sources. Service-life values beyond 20 years are power-law extrapolations and are illustrative only.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7236009170
FWCI:
Cites: 46
Cited by:
Related to: 10