Dataset and Reproducibility Materials for Machine-Learning Prediction of Bearing Capacity of Ring Foundations on Fiber-Reinforced Sandy Soil

This repository contains the datasets and supporting machine-learning outputs associated with the revised analysis of the study entitled “Interpretable Machine Learning Prediction of Bearing Capacity of Ring Foundations on Fiber-Reinforced Sandy Soil.” The complete modeling dataset contains 900 records, comprising 128 article-consistent reconstructed non-augmented records and 772 physics-constrained synthetic/augmented records. The 128 reconstructed records were divided before augmentation into 90 training, 19 validation, and 19 test records. Physics-constrained augmentation was applied exclusively to the training subset using truncated Gaussian perturbation (386 records) and Latin Hypercube Sampling (386 records), resulting in a final dataset of 862 training, 19 validation, and 19 test records. No augmented records were included in the validation or test subsets. The repository includes the complete modeling dataset, reconstructed-record dataset, dataset-split summary, reconstructed-condition summary, multicollinearity and correlation analyses, model-selection results, final test metrics, test predictions, permutation feature-importance results, and a validation report. Gradient Boosting, XGBoost, and Random Forest models were evaluated using five-fold cross-validation and an independent validation subset. Gradient Boosting was selected as the final model based on the lowest validation RMSE. The final holdout-test performance was R² = 0.999175, RMSE = 3.016 kPa, MAE = 2.356 kPa, and MAPE = 0.642%. IMPORTANT DATA-PROVENANCE NOTE: The 128 non-augmented records provided in this repository are article-consistent reconstructed records. They are not recovered row-level raw laboratory measurements and should not be interpreted, described, or cited as original raw experimental data. The additional 772 records are explicitly physics-constrained synthetic/augmented records. Accordingly, this repository is intended to support transparency, reproducibility, and interpretation of the revised machine-learning workflow within the reconstructed and physics-constrained modeling domain.