Kilometer-Scale Climate Model Matches Station-Based Products for Simulating Chile’s Water Cycle

A sweeping test of one of South America’s most ambitious climate simulations has found that a convection-permitting weather model can rival Chile’s best observation-based datasets at driving hydrological models, even though it struggles to pinpoint individual rainstorms. The study, published in Hydrology and Earth System Sciences by Sofía Segovia-Escafi of the Universidad de Chile and colleagues, evaluated the WRF-SAAG dataset — a 4-kilometer-resolution simulation of the entire continent produced by the South American Affinity Group with the Weather Research and Forecasting (WRF) model — against two widely used gridded products, CR2MET and RF-MEP, across continental Chile from April 2001 to March 2018.

The WRF-SAAG dataset was created by dynamically downscaling the ERA5 reanalysis with WRF version 4.1.5 at 4 km grid spacing and 61 vertical levels over 2000–2021, using the Thompson microphysics scheme, the RRTM radiation scheme, the YSU boundary layer, the Noah-MP land surface model, and a groundwater scheme. Because the simulation includes no data assimilation or spectral nudging, it is a free-running climate experiment: it inherits large-scale weather from ERA5 but resolves convection and mesoscale processes explicitly, without the cumulus parameterizations identified as a major error source in coarser regional climate models. That design makes it physically consistent across dozens of variables, but it was never meant to reproduce every single storm at a specific place and hour — a distinction the evaluation carefully preserves.

Chile is a brutal proving ground for any meteorological product. Continental Chile stretches more than 4,200 kilometers, from the hyper-arid Atacama at roughly 17° S to the wet, storm-lashed fjords of Patagonia near 56° S, compressed between the Pacific Ocean and the Andes Cordillera. The country’s water directorate divides this span into five macrozones: the Far North (17.5–26° S), Near North (26–32.18° S), Central Chile (32.18–36.4° S), Southern Chile (36.40–43.7° S), and Austral Chile (43.7–56° S). Spatially averaged annual precipitation ranges from 129 mm in the Far North, where summer rain arrives with the Altiplano rainy season, to roughly 2,790 mm simulated by WRF-SAAG in the Austral zone, with entirely different storm mechanisms in between. Station networks thin out dramatically toward both extremes, leaving the Far North and Patagonia with the largest disagreements among existing products — annual precipitation estimates there diverge by as much as 2,203 mm.

To benchmark the three datasets, the team drew on quality-controlled records from 494 precipitation stations and roughly 96 temperature stations maintained by Chile’s water directorate and national weather service, all passing homogeneity tests and screening for outliers. They also compared how well each dataset could replicate historically observed daily precipitation events exceeding thresholds of 1, 5, 10, and 20 mm per day, using standard contingency-table diagnostics: probability of detection, critical success index, false alarm ratio, and percentage correct. Accuracy of daily precipitation and extreme temperatures was scored with Tang’s modified Kling–Gupta efficiency, a dimensionless metric combining correlation, bias, and variability.

The results split sharply by climate. All three datasets performed best where rain is plentiful: in the Southern macrozone, median Kling–Gupta efficiencies reached 0.65 for WRF-SAAG, 0.79 for CR2MET, and 0.76 for RF-MEP during summer, and stayed at or above 0.65 year-round. In the arid north, performance collapsed — during summer in the Near North, median efficiency was negative for all three products, and 23 stations recorded no precipitation at all during the study summers, rendering the metric undefined. Winter improved matters in the Near North, where medians climbed above 0.61, but the Far North remained the weakest region, with medians of −0.14, 0.47, and 0.19 for WRF-SAAG, CR2MET, and RF-MEP respectively.

Event detection told a similar story of competing strengths. CR2MET achieved the highest probability of detection for light rain above 1 mm per day, with median values generally above 0.85 across macrozones, while RF-MEP led for moderate and heavy events above 5, 10, and 20 mm per day, achieving the highest critical success index for those thresholds. CR2MET posted the lowest false alarm ratios at the two heaviest thresholds. The authors note an important caveat: CR2MET and RF-MEP were built partly from the very stations used in this evaluation, giving them an inherent advantage at point locations, whereas WRF-SAAG faced a fully independent test under deliberately stringent event-by-event scoring that penalizes any spatiotemporal displacement of a simulated storm.

Temperature comparisons favored CR2MET overall, with median efficiencies of at least 0.71 for daily minimum temperature in Central, Southern, and Austral Chile, versus a maximum of 0.6 for WRF-SAAG in the Austral zone. WRF-SAAG ran up to 5 °C warmer than CR2MET in climatological minimum temperatures along the coast and central valley between the Far North and Central Chile, and consistently underestimated daily maximum temperatures nationwide, with median bias terms of −0.63 and −0.6 in the Far and Near North. Maximum temperature skill was better than minimum for both products — a familiar pattern, since pre-dawn minimum temperatures in weakly mixed air are strongly shaped by local surface processes that a grid cell may not represent. Still, WRF-SAAG captured temperature variability closer to observed levels, with median variability ratios of 1.00–1.04 for minimum temperature against CR2MET’s 0.88–0.94.

The most consequential test came next: feeding each dataset into a conceptual rainfall–runoff model (the TUW model, calibrated in 44 minimally disturbed catchments from the CAMELS-CL dataset) and scoring the simulated water cycle against streamflow gauges, satellite evapotranspiration, satellite soil moisture, and a snow water equivalent reanalysis. Here the free-running simulation held its own. WRF-SAAG produced streamflow performance metrics comparable to CR2MET in Central, Southern, and Austral Chile, with the best median results in the Southern macrozone, where objective function values during calibration reached 0.87–0.89 across forcing combinations. In Patagonia, the RF-MEP-based configurations performed worst, overestimating monthly runoff by amounts that could exceed 100 mm per month, while WRF-SAAG and CR2MET kept biases for key runoff signatures below 15.4 percent. Snow simulations were similarly competitive: median Kling–Gupta efficiencies of 0.61 for WRF-SAAG and 0.60 for CR2MET during calibration, though all configurations degraded markedly in the evaluation period.

WRF-SAAG did lag in two hydrological outputs. Evapotranspiration simulations driven by WRF-SAAG scored lowest across all macrozones when compared against the MOD16A2GF satellite product, and simulated soil moisture correlations, while highest for WRF-SAAG at 0.71–0.73, were evaluated only for timing rather than magnitude because the conceptual model’s soil reservoir is not physically equivalent to the satellite’s 2–5 cm surface layer. The authors stress that neither variable entered the calibration, so multi-objective calibration could plausibly close the gap. In the Far and Near North, all forcing combinations failed to reproduce annual runoff cycles, reflecting both the difficulty of modeling nearly dry basins and the limitations of a bucket-style model in regions where groundwater and snow processes dominate.

The broader takeaway is cautiously optimistic for a region starved of ground data. Despite substantial climatological differences — WRF-SAAG simulated 533 mm more annual precipitation than CR2MET in the Austral macrozone and 241 mm less in the Southern zone — the kilometer-scale model delivers daily-scale hydrological information of genuinely useful quality where station density allows, and offers physically consistent wind, humidity, pressure, and radiation fields that observational products lack, opening the door to process-based hydrological modeling and pseudo-global-warming climate impact experiments. For pinpointing precipitation extremes and temperatures at specific locations, observation-based products like CR2MET remain superior. But in the Andes, where rain gauges systematically undercatch snow and stations cluster in valleys, the study suggests that convection-permitting simulations have earned a seat at the table for water resource planning in a country living through one of the world’s most closely watched megadroughts.

Subject of Research: Evaluation of the WRF-SAAG convection-permitting climate simulation against observation-based meteorological products for hydrometeorological applications in Chile

Article Title: Benchmarking convection-permitting climate simulations for hydrometeorological characterizations: a comparative study of WRF-SAAG and observation-based products in Chile

Article References: Segovia-Escafi, S., Mendoza, P. A., Lagos-Zúñiga, M., Scaff, L., & Prein, A. F. (2026). Benchmarking convection-permitting climate simulations for hydrometeorological characterizations: a comparative study of WRF-SAAG and observation-based products in Chile. Hydrology and Earth System Sciences, 30(20), 6481-6510. https://doi.org/10.5194/hess-30-6481-2026

Image Credits: AI Generated

DOI: 10.5194/hess-30-6481-2026

Keywords: WRF-SAAG, convection-permitting climate models, Chile, hydrological modeling, precipitation, gridded meteorological datasets, CR2MET, RF-MEP, Andes, ERA5, streamflow, hydrology