West-WRF AI 2-km: High-Resolution Prediction of Integrated Vapor Transport and Precipitation

1. Introduction

• Orographic precipitation is difficult to simulate because it spans large-scale moisture transport, mesoscale turbulence, and cloud microphysics, which coarse-resolution models struggle to capture.These processes contribute to sharp spatial contrasts in rainfall and snowfall.Source: §1 · p. 2
• Regional NWP can represent fine-scale processes better but is computationally expensive and parameterization-dependent, while coarse global models may underestimate extreme precipitation during ARs.Coarse global AI models also have limited representation of extremes, complex terrain, and regional AR-related precipitation over the western U.S.Source: §1 · p. 3
• Stretched-grid models retain coarser global coverage while increasing resolution over a target region, preserving large-scale context while resolving local terrain and mesoscale processes.The introduction describes this approach as particularly suitable for AR prediction.Source: §1 · p. 3
• The study develops West-WRF AI 2-km, with approximately 2-km resolution over the western U.S. and Northeast Pacific and approximately 31-km resolution elsewhere globally.It focuses on precipitation and IVT during AR events.Source: §1 · pp. 3–4
• West-WRF AI 2-km is described as the first AI weather forecasting system evaluated specifically for AR prediction at approximately 2-km resolution over the region.The system focuses on AR-related IVT and orographic precipitation over the western U.S.Source: §1 · p. 4

• Fine-tuning uses 2012–2020 reanalysis, with winters 2020–2023 held out; benchmarks include West-WRF AI 6-km, IFS 9-km, and West-WRF 3-km/9-km.Source: a. Training Datasets · p. 4, f. Benchmark Models · pp. 7–8
• The workflow pretrains a 31-km global model on 40 years of ERA5, fine-tunes it with CW3E 2-km reanalysis, then trains with an eight-step autoregressive rollout.At inference, the model runs for up to 24 six-hour steps (144 hours).Source: b. Training Framework · p. 5
• The Anemoi Graph Transformer encodes 172 input channels on an icosahedral graph, applies processor-mesh message passing, and decodes forecasts to the latitude–longitude grid.Source: c. Stretched-grid Graph Transformer · p. 6
• Precipitation is evaluated against daily PRISM gridded observations and quality-controlled hourly CNRFC gauges, including 526 valid stations across the common model domain.Source: e. Observational Datasets · p. 7
• IVT verification uses about 2,360 AR Reconnaissance dropsonde profiles, an independent reference derived directly from observed moisture and wind profiles rather than reanalysis.Source: e. Observational Datasets · p. 7
• Evaluation covers lead times up to 120 hours and compares both the 2-km domain and its surrounding 31-km global-grid region.Source: g. Model Evaluation · p. 8

AR-Based Precipitation Events

• AR-related events use Ralph et al. (2019)’s scale and rank ≥ 2 San Francisco cases from Baño-Medina et al. (2025), spanning November 2020–March 2023.The events’ maximum IVT ranges from approximately 498–1095 kg m⁻¹ s⁻¹.Source: 1) AR-BASED PRECIPITATION EVENTS · p. 8

• Verification assesses RMSE, CC, ETS, RAPSD, FSS, POD, FAR, and CSI using forecast and observation fields.Source: 2) METRICS · p. 9

Confidence Intervals

• 90% confidence intervals use paired 33 × 33-cell spatial-block bootstrap resampling for 1,000 iterations, with block size based on PRISM’s 132-km decorrelation distance.The 5th and 95th percentiles define the intervals; the same resampled blocks are used across forecast systems.Source: 2) METRICS · p. 9

• ETS assesses event detection using a fixed NWS precipitation threshold of 0.25 mm day⁻¹ and wet-day climatological percentile thresholds P50, P90, and P98 for November–March; FSS evaluates spatial skill as a function of neighborhood size.Source: 2) METRICS · pp. 9–11
• RAPSD standardizes precipitation fields, computes a 2-D Fourier power spectrum, radially averages it, and converts wavenumber to spatial wavelength using grid spacing.Source: 2) METRICS · p. 10
• IVT event thresholds are 250, 500, and 750 kg m⁻¹ s⁻¹, while gauge-based precipitation events use each station’s 2010–2023 climatological 95th percentile.POD, FAR, and CSI assess detection, false alarms, and combined detection skill, respectively.Source: 2) METRICS · p. 11

a. Verification Against Gridded Observations

• At 4-day lead, West-WRF AI 2-km captures localized coastal and Sierra Nevada gradients, with 0.96 mm day⁻¹ RMSE—4.0% below West-WRF 3-km and 15.0% below IFS 9-km.Its RMSE is essentially on par with West-WRF AI 6-km; its domain-mean bias is +0.17 mm day⁻¹.Source: Figure 2 · p. 12, a. Verification Against Gridded… · p. 13
• At rainy-day p90 and p98, West-WRF AI 2-km reduces RMSE versus West-WRF 3-km by 7.2–12.8% and versus IFS 9-km by 11.0–17.4%.It has the closest agreement with PRISM domain means at these percentiles, though its RMSE is slightly higher than West-WRF AI 6-km.Source: Figure 2 · p. 12, a. Verification Against Gridded… · p. 14
• Against PRISM’s 38.94% rainy-day frequency, West-WRF AI 2-km overestimates frequency by 17.7%; at p50, it overestimates precipitation by 2.4% with the smallest absolute bias and highest CC.Source: Figure 2 · p. 12, a. Verification Against Gridded… · pp. 13–14
• For AR precipitation, AI 6-km gives the closest domain-mean ratio estimate and lowest RMSE, while AI 2-km has the highest spatial correlation but underestimates the domain mean by 9.7%.Source: Figure 2 · p. 12, a. Verification Against Gridded… · p. 14
• Across 1–4 days, AI forecasts maintain lower total-precipitation RMSE than West-WRF 3-km and IFS 9-km, while AI 2-km correlation exceeds IFS 9-km from day 2 onward.AI 2-km also outperforms West-WRF 3-km in correlation at 3 days or longer; IFS 9-km AR-precipitation error rises above both AI forecasts by day 4.Source: a. Verification Against Gridded… · p. 14, Figure 3 · p. 15
• At p90 and p98, AI 2-km improves ETS over West-WRF 3-km by 9–21% and 17–40%, respectively, with skill comparable to AI 6-km.At p98, its ETS improvement over IFS 9-km is 2–34%.Source: Figure 3 · p. 15, a. Verification Against Gridded… · p. 15
• Across 2–4-day leads, all models match PRISM low-to-moderate precipitation amounts closely (R² = 0.98–1.00), but AI forecasts better reproduce the upper tail than IFS 9-km while still underestimating the most extreme amounts.Source: Figure 4 · p. 16, a. Verification Against Gridded… · p. 16

Results

• At the 600-km window, West-WRF AI 2-km ranked second in FSS behind West-WRF 3-km, exceeding IFS 9-km by 11.6% at 2 days and 20.1% at 4 days.AI 2-km ranked among the two best-performing forecasts across lead times and window sizes.Source: a. Verification Against Gridded… · p. 17, Figure 5 · p. 18
• For the AR Scale 2 case, West-WRF AI 2-km had the lowest spatial RMSE at 5.52 mm, 33.3% below IFS 9-km and 18.7% below West-WRF 3-km.It captured the narrow coastal band and localized maxima with strong spatial agreement to PRISM despite a domain-mean overestimate of 7.5%.Source: a. Verification Against Gridded… · pp. 18–19, Figure 6 · p. 19
• For the AR Scale 3 case, West-WRF AI 2-km’s spatial RMSE was 9.64 mm, approximately 28.0% below the 6-km forecast’s 13.38 mm.The 2-km forecast resolved both localized high-intensity cores with sharper gradients and peak magnitudes closer to PRISM, whereas the 6-km forecast smoothed them.Source: a. Verification Against Gridded… · pp. 19–20, Figure 7 · p. 20

b. Verification Against Gauge-Based Observations

• For 6-hour accumulated precipitation forecasts verified against gauges, both West-WRF AI forecasts retained CC ≈ 0.61 at 120 hours, versus 0.53 for IFS 9-km; RMSE rises and CC falls with lead time.Source: b. Verification Against Gauge-Based… · p. 21, Figure 8 · p. 22
• At 24 hours, gauge-based POD for 95th-percentile precipitation events was highest for West-WRF AI 2-km at 0.516, and CSI was 0.324 for AI 2-km and 0.320 for AI 6-km.At 120 hours, CSI was 0.186 for AI 6-km and 0.135 for AI 2-km, versus 0.105 for IFS 9-km and 0.084 for West-WRF 3-km; the AI forecasts’ FAR performance traded off by lead time.Source: b. Verification Against Gauge-Based… · pp. 21–22

c. AR Reconnaissance Dropsonde IVT

• Dropsonde-verified IVT RMSE increased with lead time for the evaluated forecasts.The passage identifies 2,360 AR Reconnaissance dropsonde samples and lead times of 24–120 hours.Source: c. AR Reconnaissance Dropsonde IVT · p. 23

3. Results

• At 120 hours, West-WRF AI 2-km/31-km RMSE was 183.30 kg m⁻¹ s⁻¹, 5.6% below AI 6-km and 8.1% below West-WRF 9-km.The corresponding CC values were 0.67, 0.64, and 0.62, respectively.Source: c. AR Reconnaissance Dropsonde IVT · p. 23, Figure 9 · p. 24
• West-WRF AI 2-km/31-km had the lowest RMSE and highest CC at every lead time, with its advantage growing at 96 and 120 hours.West-WRF AI 6-km performed between it and West-WRF 9-km; differences were small at short lead times.Source: c. AR Reconnaissance Dropsonde IVT · p. 23, Figure 9 · p. 24
• At 750 kg m⁻¹ s⁻¹, POD fell from approximately 0.75–0.76 at 24 hours to 0.39–0.40 at 120 hours, with West-WRF 9-km declining more at higher thresholds.Categorical performance degraded with lead time for all configurations; the two AI models maintained broadly comparable POD across most lead times.Source: c. AR Reconnaissance Dropsonde IVT · p. 23, Figure 9 · p. 24
• At 750 kg m⁻¹ s⁻¹ and 120 hours, FAR was about 0.49 for West-WRF AI 2-km/31-km versus approximately 0.56 for the other configurations, a 13% reduction.Source: c. AR Reconnaissance Dropsonde IVT · p. 23, Figure 9 · p. 24
• At 750 kg m⁻¹ s⁻¹ and 120 hours, CSI was 0.29 for AI 2-km/31-km, versus 0.26 for AI 6-km and 0.25 for West-WRF 9-km.These corresponded to improvements of approximately 12% and 16%, respectively.Source: Figure 9 · p. 24, c. AR Reconnaissance Dropsonde IVT · p. 25
• At 750 kg m⁻¹ s⁻¹, ETS fell from near 0.61 at 24 hours to approximately 0.27 for AI 2-km/31-km and 0.23 for both other models at 120 hours.All models retained positive ETS; West-WRF 9-km was consistently lowest, particularly for stronger events and longer lead times.Source: Figure 9 · p. 24, c. AR Reconnaissance Dropsonde IVT · p. 25

4. Discussion and Conclusions

• The 2-km configuration better represents upper-percentile precipitation intensity than 6-km, without the stronger upper-tail overestimation seen in West-WRF 3-km.Differences between the AI configurations are small near the median and grow at upper percentiles.Source: §4 · pp. 25–26
• For the December 31, 2022 case, 2-km sharpened localized cores and had lower spatial RMSE than 6-km, which smoothed and displaced the most intense core.The reported spatial RMSEs were 9.64 mm for 2-km and 13.38 mm for 6-km.Source: §4 · p. 26
• Gauge verification found AI forecasts competitive with or better than dynamical forecasts across much of the five-day period, but neither AI configuration was uniformly best for station-specific extremes.The 120-hour reversal involves rare station-specific 95th-percentile events with limited effective sample size.Source: §4 · pp. 26–27
• The configurations differ in both regional resolution and outer-domain setup, so controlled experiments are needed to isolate the effect of inner-grid refinement.Reanalysis initialization may advantage West-WRF AI relative to dynamical forecasts initialized from operational analyses; operational-analysis testing would better represent real-time performance.Source: §4 · pp. 26–27
• At short IVT lead times, lower-percentile precipitation intensity, and large-neighborhood skill, 2-km performs comparably to—not clearly better than—6-km.The value of refinement depends on forecast variable, spatial scale, lead time, verification method, and application.Source: §4 · pp. 26, 28
• Pointwise verification can penalize displaced sharp features twice, while neighborhood verification tolerates small displacements; MSE training favors spatially averaged predictions under positional uncertainty.The displacement sensitivity of sharper 2-km features grows with lead time.Source: §4 · pp. 26–27
• The study is a deterministic baseline; the 2-km refinement and fine-tuning are limited to the CW3E reanalysis domain, and ERA5 training versus operational-analysis initialization remains unevaluated.The authors identify ensemble and probabilistic prediction, wider 2-km coverage, and operational-analysis initialization as future directions.Source: §4 · pp. 27–28
• The clearest 2-km benefits are for localized, terrain-sensitive extreme AR precipitation and intense IVT, while large-scale skill shows no evident reduction.The two configurations remain comparable for domain-mean quantities, broader spatial scales, lower precipitation thresholds, and short-lead IVT.Source: §4 · p. 28