CAMELS-US Hydrological Model Intercomparison Project (MIP) — ERA-Physics-Guided-PUR (10-model ensemble) Daily Streamflow Hydrographs (531 Basins, Water Years 1990–1999)

This repository contains the complete 10-water-year daily streamflow simulation archive (1989-10-01 through 1999-09-30; 3,652 daily steps across 531 canonical CAMELS-US basins) for ERA-Physics-Guided-PUR (10-model ensemble), evaluated as part of the CAMELS-US Hydrological Model Intercomparison Project (MIP) (grey-nearing/camels-us-mip). Model Name: ERA-Physics-Guided-PUR (10-model ensemble) Target Leaderboard: pur__multi_camels Spatial Evaluation Regime: Prediction in Ungauged Regions (PUR — Canonical 18-Region Leave-One-HUC02-Out Cross-Validation) Meteorological Forcings: Multi-Forcing (Daymet + Maurer + NLDAS) Ensemble Size: n = 10 independent random seed realizations (seeds = 0..9) Open-Source Model Code: https://github.com/grey-nearing/camels-us-mip/tree/main/models/era_pur_physics_guided ERA Discovery Notebook: https://github.com/grey-nearing/camels-us-mip/blob/main/era_experiments/era_pur_physics_informed_winner.ipynb Model Architecture & Experimental Protocol Architecture: Physics-Guided Regional Extrapolation Architecture (`PURPhysicsModel`) combining continuous degree-day snowpack/melt state tracking, Priestley-Taylor / Hamon potential evapotranspiration and rolling catchment moisture deficit features, low-rank static attribute FiLM bottleneck modulation, and multi-timescale recurrent routing across Daymet, Maurer, and NLDAS. Strict Non-Leakage Protocol: Architecture discovered via automated Empirical Research Agent (ERA) tree search (XID 302979341, Winner: Node 206) using a strictly held-out regional validation split (HUC-02 Region '01' New England held out as the ungauged validation region; remaining 17 HUC-02 regions trained on 1999–2008) with zero access to the 1989–1999 Newman evaluation period. For benchmark evaluation, 10 independent seeds (seeds 0–9) were trained across all 18 canonical Leave-One-HUC02-Out PUR folds (180 models total) on the Newman training window (1999-10-01 to 2008-09-30) and evaluated out-of-sample on each held-out HUC-02 region over 1989-10-01 to 1999-09-30. Benchmark Performance Summary (531 CAMELS-US Basins, 1989-10-01 to 1999-09-30) Ensemble Median NSE (n = 10): 0.684 Single-Model Median NSE (mean ± SD across 10 seeds): 0.653 ± 0.007 Ensemble Mean NSE: 0.524 Ensemble Median KGE: 0.657 Ensemble Median Peak-Flow Bias |FHV|: 23.40% Catastrophic Failure Basins (NSE ≤ 0): 28 / 531 Included Files era_pur_physics_guided_pur_18huc02_multi_camels_hydrographs_531x3652.npz: Compressed NumPy archive containing qsim / qsim_members of shape (10, 531, 3652) (float32, mm/day), qsim_ensemble_mean of shape (531, 3652), qobs of shape (531, 3652), basins of shape (531,), dates of shape (3652,), and seeds of shape (10,). era_pur_physics_guided_pur_18huc02_multi_camels.zarr.zip: Consolidated Zarr v2 store containing q_sim (10, 531, 3652), q_obs (531, 3652), basin (531,), time (3652,), member (10,), and all 14 pre-computed per-basin hydrological evaluation metrics defined by Gauch et al. (2023). README.md: Complete documentation, schema reference, and Python loading example.