Data and computational materials for a state variable strength model of frozen sand–gravel mixtures

This repository provides the data and computational materials supporting the manuscript “A state variable strength model for frozen sand–gravel mixtures based on the partitioning of load-bearing systems.” The study is based on 108 unconfined compression tests on frozen sand–gravel mixtures under different gravel contents, degrees of saturation, relative densities, strain rates, and temperatures. XGBoost interpreted using Shapley additive explanations (SHAP) was used to identify variable contribution patterns and conditional relationships. These results, together with prior knowledge of frozen soil mechanics, were used to construct physically interpretable state variables and a strength model based on the partitioning of sand–ice, sand–ice–gravel, and gravel–ice load-bearing systems. The repository contains six structured Excel workbooks and seven Jupyter notebooks. The workbooks provide the experimental datasets, XGBoost–SHAP results, parameter bounds and identified parameters, specimen-level predictions and strength decomposition, repeated-partition stability results, and recalibration results for three independent frozen soil datasets. The notebooks implement the corresponding workflows, including XGBoost training and SHAP analysis, Sobol candidate sampling, XGBoost surrogate validation, particle swarm optimization, assessment across 100 random training/test partitions, and dataset-specific recalibration. The original stress–strain histories are not included. Reproduction begins from the processed specimen-level test conditions and unconfined compressive strength values. The repository supports verification of the reported model predictions, parameter identification, strength decomposition, partition stability, and recalibration across datasets.