Reproducibility of Machine Learning Accuracy for Chemically Stabilised Soils using an Open UCS Database

Machine learning models for the unconfined compressive strength (UCS) of lime and cement stabilised soils often report test coefficients of determination (R 2 ) above 0.9. Cross-validation is widely used to tune these models, but the headline score is usually taken from one train-test split, and its sensitivity to that split is rarely reported. This study examines that sensitivity on an openly licensed database of 194 records. A single 80:20 split was repeated 200 times with fixed hyperparameters, and five model configurations were compared under ten repeats of five-fold nested cross-validation, together with monotonic binder constraints, empirical interval coverage and a leverage-based applicability domain. Across the 200 splits, test R 2 ranged from −0.235 to 0.962 (mean 0.782, SD 0.181). Under nested resampling a regularised multilayer perceptron (R 2 = 0.880 ± 0.103) outperformed extreme gradient boosting (0.810 ± 0.195), reversing the single-split ranking. Monotonic constraints removed physically inadmissible binder responses with no measurable loss in R 2 or RMSE and an MAE increase of 0.007 MPa. Naive bootstrap intervals covered 33.3% of observations against a nominal 90%, rising to 87.2% with out-of-fold residuals, and five records fell outside the applicability domain. Model interpretation agreed with the original study despite the differences in accuracy. Reporting accuracy as a distribution, with coverage and domain checks, is recommended for small geotechnical databases.