Data and Code Supplement to "Machine learning metamodelling for global sensitivity analysis"

Description This data and code supplement provides the material required to reproduce the numerical experiments and figures presented in the manuscript “Machine learning metamodelling for global sensitivity analysis”. The repository contains Monte Carlo simulations, Sobol’ GSA results, machine learning (ML) metamodel outputs, and the scripts used to compute sensitivity measures and generate all visualizations. The data were produced using three conceptual hydrologic models: Hydrologiska Byråns Vattenbalansavdelning (HBV), HyMod, and Variable Infiltration Capacity (VIC) version 5. For each model–catchment combination, parameter samples were generated using Sobol’ quasi-random sequences and were used to compute the Kling–Gupta Efficiency (KGE), which serves as the model output of interest throughout the study. The repository includes the KGE outputs of these Monte Carlo experiments for each model, which form the basis for both the variance-based global sensitivity analysis and the training of the ML metamodels. The repository includes computed Sobol first-order and total-effect indices (S_i and T_i , respectively), as well as feature importance measures derived from random forest (RF), neural network (NN), and linear model (LM) surrogates. In particular, Permutation Variable Importance (PVI_i) and SHapley Additive exPlanations (SHAP_i) are provided. For the LM metamodel, the standardized regression coefficients β_i and β_i^2 are also available. All analyses were implemented in Python. The scripts contained in the repository reproduce the computation of sensitivity and feature importance measures and the generation of the figures presented in the manuscript. The workflow is fully reproducible from the supplied data and does not require additional hydrologic model simulations. This supplement is intended to facilitate transparency, reproducibility, and reuse of the proposed framework. Researchers may use the material to benchmark alternative surrogate models, evaluate additional sensitivity metrics, or extend the analysis to different modelling contexts.