HRES: High-Res Weather Forecast

• HRES is a high-resolution deterministic weather forecasting system from ECMWF that provides detailed, physically consistent predictions.
• It employs a 0.1° horizontal resolution and advanced data assimilation to accurately capture atmospheric states and extreme weather events.
• HRES sets the benchmark by outperforming AI models in forecasting record-breaking events, ensuring robust decision support for critical applications.

The High RESolution Forecast (HRES) refers to the operational high-resolution deterministic numerical weather prediction (NWP) system developed and maintained by the European Centre for Medium-Range Weather Forecasts (ECMWF). HRES is widely recognized as the “gold standard” in global deterministic weather forecasting, employing cutting-edge physics, advanced data assimilation, and a fine spatial and temporal resolution. Its consistent skill, especially for extremes and out-of-sample situations, makes it a benchmark for both physical NWP and emerging AI-based forecasting systems.

1. Definition and Role within Numerical Weather Prediction

HRES is the high-resolution, deterministic member of ECMWF’s global weather forecast suite. As of recent cycles, it runs at a horizontal resolution of approximately 0.1° (about 9 km at the equator), with 137 vertical levels, and produces forecasts four times per day out to 10 days and twice daily to 15 days. HRES is complemented by the ECMWF ensemble prediction system (ENS) and reforecast systems. Its core objective is to provide highly detailed, physically consistent, and skillful predictions of atmospheric state variables (e.g., temperature, pressure, humidity, wind) necessary for weather hazard forecasting, operational decision support, climate monitoring, and as input or boundary conditions for limited-area and sectoral models.

HRES solves the nonhydrostatic primitive equations of atmospheric motion using a semi-Lagrangian, semi-implicit spectral–finite difference dynamical core. Data assimilation integrates a diverse set of global observations—including satellites, radiosondes, and surface stations—via 4D-Var or hybrid-ensemble variational techniques.

2. Verification, Skill, and Performance Metrics

HRES is evaluated against both ground truth observations and reanalysis datasets using rigorous forecast verification metrics. Commonly used measures include:

• Root Mean Squared Error (RMSE) of gridded fields, weighted by the latitude-dependent gridcell area: $$\text{RMSE}(\tau) = \sqrt{\frac{1}{\sum_{(s,t_0)\in I} \omega_{s}}\sum_{(s,t_0)\in I} \omega_{s} \left(\hat{x}_{s,t_0}^\tau - x_{s,t_0+\tau}\right)^2}$$ where events may be filtered for all, extreme, or record-breaking conditions, and $\omega_{s}$ denotes the gridcell weight (Zhang et al., 21 Aug 2025).
• Anomaly Correlation Coefficient (ACC), mean absolute error (MAE), standard deviation of error (SDE), and Continuous Ranked Probability Score (CRPS), among others (Bremnes et al., 2023, Chen et al., 2023, Sun et al., 10 Aug 2024).
• Specialized skill assessment for specific application variables (e.g., precipitation, wind, tropical cyclones) using binary event skill scores, precision-recall for records, and maxima/minima analysis.

Verification results demonstrate that HRES exhibits robust skill across standard variables and regimes:

• For general deterministic scores (RMSE, MAE), HRES outperforms AI-based and low-resolution models, especially on wind speed and in capturing the amplitude of local weather events (Bremnes et al., 2023, Baran et al., 18 Jun 2025).
• Critically, HRES outperforms even state-of-the-art AI models (e.g., GraphCast, Pangu-Weather, Fuxi) on record-breaking extremes—both in RMSE and event detection precision-recall curves. AI models tend to underpredict the intensity and frequency of such events, particularly as the event magnitude exceeds the training distribution maximum; the HRES error remains nearly constant with increasing record exceedance, while that of AI models increases almost linearly (Zhang et al., 21 Aug 2025).

3. Comparison with AI and Data-Driven Forecasting

While machine learning–based global weather models increasingly match or exceed the average skill of HRES for “typical” extreme events, HRES’s physical basis allows it to extrapolate reliably to unprecedented atmospheric regimes. Several recent studies reveal (Zhang et al., 21 Aug 2025):

• For record-breaking 2-meter temperature, 10-meter wind, and related extremes, HRES demonstrates significantly lower forecast errors and systematically more accurate event forecasts across short and medium lead times, compared to AI models.
• AI models exhibit implicit capping generated by their training sample maxima, underestimating new record events ("underprediction of intensity") and failing to predict the correct number of records ("underprediction of frequency"). HRES, by explicitly solving governing equations, does not share this limitation.
• In verification over compound metrics (e.g., heat index, wind chill), HRES maintains more accurate structural error patterns (improvement as forecast time converges with event time), while AI models often display persistent biases even at short lead times (Pasche et al., 26 Apr 2024).
• For wind and power sector–relevant forecasts, post-processing (e.g., via EMOS) can help calibrate and mitigate the resolution gap, but pure high-resolution ensembles (HRES members) remain superior in skill over their low-resolution or mixed counterparts (Baran et al., 18 Jun 2025).

4. Methodological Innovations and Operational Considerations

HRES continues to adopt and drive innovations in the field:

• Dual-resolution and ensemble techniques explored in operational deployment leverage complementary high- and low-resolution runs to balance computational cost and forecast spread.
• Statistical and deep learning–based post-processing (for example, the use of Bernstein Quantile Networks or ensemble model output statistics) is routinely used operationally to further calibrate HRES forecasts and account for deficiencies in raw outputs (Bremnes et al., 2023, Baran et al., 18 Jun 2025).
• For ocean and coupled atmosphere–ocean applications, HRES data are increasingly integrated with both deterministic and probabilistic ensemble systems to provide seamless Earth-system forecasting (Garcia et al., 21 Jan 2025).
• HRES outputs serve as essential boundary and initial conditions for regional and limited-area models, and as a global reference for downscaling, hydrological, and renewable energy applications (Xu et al., 29 Jan 2024, Wang et al., 7 May 2025).

In terms of computational efficiency, HRES demands high-performance computing resources (large HPC clusters), which remains a consideration versus the relatively lower cost of AI inference in operational settings. However, for societal stakes—such as early warning, disaster management, and infrastructure protection—its physical consistency is preferred due to its reliability for out-of-sample extremes.

5. Limitations, Ongoing Challenges, and Controversies

The most critical limitation of HRES, compared to emerging AI models, is computational cost—high-resolution, global, deterministic NWP models require significant supercomputing resources for timely execution. This motivates research into post-processing, hybrid statistical–physical modeling, and the integration of AI to accelerate and supplement high-fidelity NWP outputs.

Whereas AI models continue to advance in operational realism and speed, current evidence suggests that, for high-stakes situations involving record-breaking extremes, sole reliance on AI models remains inadvisable due to their inability to reliably extrapolate beyond the historical record envelope (Zhang et al., 21 Aug 2025). A plausible implication is that high-resolution physically based models like HRES will remain essential in critical operational workflows for the foreseeable future.

6. Implications for Future Model Development

The continued superior skill of HRES on record-breaking extremes and its operational robustness direct the research and operational communities toward hybrid model developments:

• AI–NWP fusion approaches seek to combine the fast inference and flexible calibration of AI with the physical extrapolation and structural constraint of NWP.
• Detailed, event-driven verification, as well as inclusion of comprehensive variable sets (e.g., surface humidity for heat-health impacts), are essential both for public trust and scientific rigor.
• Despite the computational burden, improvements in parallelization, data assimilation, and joint atmosphere–ocean–land coupling are expected to further extend HRES’s forecast horizon and skill ceiling.

Researchers are therefore encouraged to use HRES not only as a baseline for evaluating new deep learning forecasting systems, but also as a core component of operational early warning and impact assessment frameworks, particularly in the context of a warming climate and increasing frequency of out-of-distribution extremes.

Table: RMSE Comparison for Record-Breaking Extremes (Excerpt from (Zhang et al., 21 Aug 2025))

Forecast System	Typical Extremes: RMSE	Record-Breaking Extremes: RMSE
HRES (ECMWF)	~Equal or slightly higher	Lower than all AI models
Best AI Model (e.g., GC)	Best or equal to HRES	Significantly higher than HRES

This table demonstrates that while advanced AI models can now match or marginally outperform HRES in aggregate or “moderate” extreme event skill, HRES remains distinctly superior on out-of-sample, record-breaking forecasts. The physical constraints of the HRES NWP system confer reliable extrapolation, a critical shortcoming in all current AI models.

---

In summary, the High RESolution Forecast (HRES) represents the current apex of operational deterministic physics-based global weather prediction. HRES’s capacity for accurate simulation, especially of record-breaking extremes, continues to define the benchmark for both forecast skill and methodological innovation, even as machine learning systems accelerate progress in complementary domains. These findings suggest that responsible operational and research forecasting for the foreseeable future demands the integration of HRES or similar physical models as a foundational component.