Beyond Architecture: Optimization and Hydrologic Regime Reveal Performance Headroom in Deep-Learning Rainfall–Runoff Modelling

This repository provides the research data, model outputs, analysis scripts, and supporting materials associated with the study “Beyond Architecture: Optimization and Hydrologic Regime Reveal Performance Headroom in Deep-Learning Rainfall–Runoff Modelling.” The study investigates how hyperparameter optimization, hydrological regime information, and ensemble design influence the predictive performance of deep-learning rainfall–runoff models across 531 CAMELS-US catchments. The repository supports the evaluation of regional modelling strategies, hydrologically informed ensembles, configuration-level contributions, and sensitivity analyses. Contents of this release Original research materials: Ten files retained from earlier versions of the dataset, including model archives, predictions, metrics, catchment attributes, and supporting data. 2026 revision analyses: CAMELS_US_Revision_2026_Analysis_Data_v3_additions.zip, containing the additional analysis materials and results underlying the current study, including Top-10 and hydrologically informed ensemble diagnostics, Top-18 sensitivity comparisons, and supporting figure and table resources. Additional Top-18 experiments: CAMELS_US_Top18_Additional8_Raw_10Seed_v3.zip, containing the archived outputs and supporting experiment artifacts for eight additional model configurations evaluated over ten random seeds (80 runs). Version and provenance note This is version 3 of the CAMELS-US research deposit. The files inherited from earlier dataset releases are retained unchanged for provenance and continuity, while the two additional archives document the analyses and experiments introduced in the 2026 release. The version history of the deposit is maintained through Zenodo.