Source-aware machine learning for stabilised expansive soils

Machine-learning models for stabilised expansive soils are often assessed by random record splitting, which may place observations from the same experimental source in both training and validation sets. This study proposes a source-aware framework using 1728 unconfined compressive strength (UCS) observations from 24 original-study groups and 476 California bearing ratio (CBR) observations from 17 groups. The framework combines feature-coverage checks, internal and grouped-source validation, threshold sensitivity, uncertainty reporting, regression benchmarks, training-only synthetic augmentation and emissions-based Pareto screening. Internal balanced accuracy was 0.937 for UCS screening and 0.883 for CBR screening, but decreased to 0.630 and 0.694, respectively, when complete original-study groups were held out. CBR source groups remained uneven in size and class composition, and one grouped fold was single-class at the 8% threshold; CBR transfer estimates are therefore treated as screening-level evidence. Continuous grouped-source regression remained weak (R² = 0.130 for UCS and −0.127 for CBR). Synthetic augmentation produced only small, task-dependent changes. Conservative shortlisting uses source-aware classification and source-wise ranking within observed data coverage, removes Pareto-dominated mixtures using measured gain and approximate emissions and sends retained candidates for project-specific laboratory confirmation rather than replacing UCS or CBR testing.