Single-step and multi-step forecast evaluation using NWP

To evaluate the single-step simulation performance of the AI-DUST model, we initialized it with dust emissions and concentrations from a traditional numerical model at a given time step. The same meteorological fields were used to drive both models for predicting the concentration changes in the subsequent time step (Fig. 1).

Fig. 1: Comparison between observed and AI-DUST-predicted dust concentrations at the next time step.

figure 1

Observations are derived from dust event samples in 2024 across multiple vertical levels: a–d surface, 950 hPa, 850 hPa, and 500 hPa, respectively, for grid cells without dust emissions; e–h same levels for grid cells with active dust emissions. All data are in log-transformed values.

The results show that the AI-DUST model can effectively replicate the dust diffusion, transport, and deposition processes simulated in the traditional model. Dust concentrations showed an R exceeding 0.99. Notably, performance was higher in non-emission regions compared to emission-active regions, likely due to the absence of uncertainties associated with the vertical lofting of dust into the free atmosphere. Predictions at lower atmospheric levels outperformed those at higher altitudes, as near-surface dust concentrations are typically higher and more dynamically variable, providing richer training signals for the AI model. For instance, in non-emission regions, the R for near-surface dust concentrations reached 0.9994, with an RMSE of 0.0444. In contrast, in emission-active regions at 500 hPa, the R was 0.9927, with an RMSE of 0.1068. Despite these differences, overall performance across various regions and altitudes remained robust and consistent, indicating that the model effectively captures the dominant processes in diverse conditions.

To evaluate the long-term forecasting capability of the AI-DUST model, we conducted multi-step predictions up to 80-time steps ahead (Fig. 2). While performance gradually degraded with increasing forecast lead time, the model maintained physically consistent and skillful predictions throughout the simulation period. Even at the 80th step, the R remained above 0.6. In emission-active regions, the R reached 0.71 at the final step, with an RE of −5.9%. In non-emission regions, the R was 0.63, with a mean RE of −8.3%. The relatively weaker skill in non-emission regions is primarily attributed to the cumulative propagation of errors in dust transport, deposition, and diffusion processes, which are not reset by local emissions. Conversely, in emission-active regions, ongoing dust emissions act as a recurring source that partially offsets the accumulation of transport and deposition-related errors. Each step’s emission contribution resets the error accumulation, leading to relatively stable long-term forecasts despite initial higher one-step errors.

Fig. 2: Temporal evolution of R (blue lines) and RE (gray area) between AI-DUST and the reference numerical model across forecast steps.

figure 2

a Surface dust concentrations in regions without dust emissions; b Surface dust concentrations in regions with active dust emissions.

Overall, the AI-DUST model demonstrates robust performance in both one-step and multi-step simulations. The results indicate strong model skill in reproducing dust concentration changes in a traditional numerical model under both emission-influenced and advection-dominated conditions.

Real-time forecast for SDS weather in East Asia 2025

Since 1 March 2025, we have been conducting operational SDS forecasts over East Asia using the AI-DUST model driven by AIFS from ECMWF, providing 10-day-ahead predictions. Model performance was evaluated against observational records during the main dust season (March–May 2025).

The model achieves a TS of 0.33 for 1-day ahead SDS event forecasts (Fig. 3). For days 2–3, the TS remains high at 0.28, gradually decreasing thereafter but staying above 0.22 throughout the entire 10-day forecast window. Notably, these TS outperform the typical performance of NWP operating in the WMO SDS-WAS Asian regional center. Most traditional models achieve TS values between 0.20 and 0.29 for 1–3 day forecasts22, indicating that the 10th-day forecast from AI-DUST performs comparably to, or even surpasses, the 3rd-day forecast of some traditional numerical models.

Fig. 3: Daily TS of the AI-DUST model system for SDS weather forecasting over East Asia across 1–10 day forecast lead times.

figure 3

The red numbers indicate the average TS for each lead day.

To further evaluate the performance of the AI-DUST model, we conducted an independent validation using ground-based PM₁₀ observations from the national air quality monitoring network. Figure 4 presents the spatial distribution of the daily R and RMSE between predicted and observed PM₁₀ concentrations over the 1–10 day forecast lead times. On day 1, the model achieves a domain-averaged spatial R of 0.57, with an RMSE of 289 μg/m³. The R remains above 0.35 for lead times up to day 5, and even at day 10, it stays above 0.30, indicating sustained skill in capturing the spatiotemporal evolution of dust pollution. The RMSE gradually increases with forecast horizon but remains below 390 μg/m³ on average across the domain. These results demonstrate improved predictive accuracy compared to traditional numerical dust models11.

Fig. 4: Spatial R and RMSE between AI-DUST predicted and observed PM₁₀ concentrations over East Asia.

figure 4

a Spatial R and b RMSE. Metrics are averaged over the study region and shown for each forecast lead time from day 1 to day 10.

During spring 2025, a total of 14 widespread SDS events occurred, all of which were used to evaluate the forecasting performance of the AI-DUST model. To benchmark its forecast skill in providing early warnings, we compared the TS of AI-DUST against those from the KMA model, a widely recognized numerical model known for its stability and accuracy in East Asia.

The results show that the AI-DUST model outperformed the KMA model in the majority of these events. The AI-DUST achieved an average TS of 0.42 over the 24–48 h forecast window, representing a 27% improvement relative to KMA (0.33). This significant enhancement demonstrates the competitive advantage of the AI-DUST model over traditional numerical models in regional SDS events forecasting.

Furthermore, as shown in Fig. 5, the TS for 9 out of the 14 events exceeded the seasonal mean (0.28) (Fig. 3), indicating consistently strong performance across strong SDS events. The model exhibited particularly high skill in strong SDS events capturing, highlighting its reliability for operational early warning systems.

Fig. 5: Comparison of TS between the AI-DUST and KMA models for 14 SDS events during spring 2025.

figure 5

The forecast data are evaluated at the 24–48-h lead time.

Forecast for strong and severe SDS events in East Asia 2025

We conducted a more detailed analysis of the forecasting performance for four significant SDS events that reached central and eastern China (Fig. 6). These events were selected due to their large spatial extent and high intensity. In the source regions of northwest China, the model showed strong temporal correlation with observations, with R between predicted and observed PM₁₀ concentrations exceeding 0.6 at most stations. RE were mostly within ±50%, indicating that the model reasonably captures dust emission dynamics in these areas. In downstream regions, central and eastern China, where dust is transported from the source areas, temporal correlations remained above 0.4, with RE ranging from 0 to 50%, suggesting a moderate underestimation of PM₁₀ concentrations during long-range transport. This is consistent with the analysis presented in Section 3.1.

Fig. 6: Evaluation of the AI-DUST model’s performance for four major dust events during spring 2025.

figure 6

a–d show the temporal R of PM₁₀ concentrations at different stations. e–h present the RE in peak PM₁₀ concentration forecasts. i–l display the spatial R for each event.

Comparisons with satellite observations (Fig. 7) further demonstrate that AI-DUST accurately reproduces the spatial extent and location of the SDS events. At the 48-h forecast lead time, the model achieved a spatial R of 0.63 for these four events, surpassing both the seasonal mean spatial correlation for 2-day forecasts (0.47) and even the seasonal mean for 1-day forecasts (0.57; Fig. 4a). This highlights the model’s enhanced capability in simulating the spatial evolution of severe SDS events, particularly under extreme conditions.

Fig. 7: Comparison between satellite-observed AOT and AI-DUST predicted dust concentration.

figure 7

show the AOT data, while b, d display the AI-DUST predictions at coincident times on 25 March (a, b) and 11 April (c, d) 2025. Satellite observations were acquired at local times of 11:00, 12:00, 14:00, 15:00, and 16:00. The AI-DUST model forecasts are sampled at the nearest forecast time step for each grid cell to match the satellite overpass time.

Notably, the SDS event from April 10 to 14 was the first in nearly a decade to reach southern China, making it one of the most unusual dust events in recent decades. Despite the absence of similar events in the model’s training data, AI-DUST successfully predicted both the spatial extent and intensity of this extreme episode. For this event, the model achieved a domain-averaged TS of 0.75. Over half of the observational stations in the affected region exhibited temporal R above 0.8 between predicted and observed PM₁₀ concentrations, with RE confined to ±50% (Fig. 6c, g). In southern China specifically, PM₁₀ concentration RE remained within ±20%, indicating high accuracy and low systematic error in this atypical region.

This successful prediction underscores the model’s strong generalization capability, particularly for rare and extreme dust storms. It highlights the potential of AI-DUST for operational forecasting of unprecedented SDS events.

Model sensitive to meteorology fields and dust emissions

We designed a set of sensitivity experiments (Supplementary Note 2) to evaluate the performance of AI-DUST in simulating dust events during spring 2025. The model was driven by different meteorological forecast fields and surface characteristics over dust source regions to demonstrate its ability to represent key atmospheric physical processes of dust particles and to understand the reasons behind the improved forecast skill observed in our system. In addition to the control run driven by AIFS, we performed parallel simulations using meteorological inputs from the National Centers for Environmental Prediction Global Forecast System (NCEP-GFS) (downscaled by WRF) and the CMA global forecast system (CMA-GFS) to assess the impact of driving meteorology on SDS forecasts.

As shown in Fig. 8, the TS of the three experiments exhibits clear differences. On day 1, the average TS is 0.33, with the lowest value at 0.22. By day 10, the average TS range from 0.23 to 0.18. To quantify the divergence among the simulations, we computed the RMSE of dust concentrations between the NCEP-driven and CMA-driven runs relative to the AIFS-driven reference. The RMSE exceeds 100 µg/m³ and increases with forecast duration, consistent with the behavior observed in traditional multi-model dust forecasting systems11. These results confirm that our model effectively captures the influence of different meteorological fields drivers on dust prediction (Fig. 9).

Fig. 8: Evaluation of AI-DUST dust forecasts driven by different meteorological fields.

figure 8

The boxplots show the TS across forecast lead times for AI-DUST simulations driven by NCEP (blue), CMA (orange), and AIFS (green). The line plots display the RMSE of dust concentrations from the NCEP-driven (blue) and CMA-driven (orange) simulations relative to the AIFS-driven reference.

Fig. 9: TS using different surface roughness density datasets.

figure 9

Simulations without high-resolution surface roughness density are shown in brown, while those utilizing enhanced data are depicted in green.

We also conducted experiments using different surface roughness densities in dust source regions. Results show that employing high-resolution surface roughness densities (Fig. 10) significantly improves the model’s forecasting skill, with an average 18% increase in TS over forecast days 1–10. The enhancement is consistent across all lead times, indicating that refined surface data consistently benefits prediction accuracy. This also demonstrates that the AI-DUST model is sensitive to dust emission processes.

Fig. 10: Architecture of the deep graph network for next time dust concentration forecast.

figure 10

The model operates on a 15 × 15 grid graph with four-connected nodes. Stacked GAT layers aggregate neighborhood information, with feature fusion from advection, emission, and initial concentration introduced after the first layer. Final predictions are generated via feature concatenation, a fully connected layer, and Sigmoid activation. Physical consistency is enforced through prediction limitation and mass conservation-aware loss. Separate models are trained for each layer, treating the concentration at one specific layer as the prediction target at a time. Training is further stratified by surface emission activity (emission vs. non-emission zones).

The sensitivity experiments demonstrate that the AI-DUST model effectively simulates the atmospheric physical processes of dust particles and exhibits clear sensitivity to both meteorological forcing fields and dust emissions. Among all experimental configurations, the highest TS are consistently achieved when the model is driven by AIFS meteorological forecasts and high-resolution surface characteristics. The superior performance of AI-DUST can thus be attributed to three key factors: its ability to accurately represent dust-related atmospheric physics, the high quality of AIFS meteorological forecasts, and the use of high-resolution surface roughness density that better constrains dust emission processes.

SDS Forecasts over North Africa and the Arabian Peninsula

To evaluate the geographical generalization capability of the AI-DUST model beyond its training domain in East Asia, we conducted zero-shot forecasting experiments for SDS events in North Africa and the Arabian Peninsula. Figure 11 compares AOT from VIIRS Deep Blue with AI-DUST predictions for 22 March and 12 May 2025.

Fig. 11: Comparison of satellite-observed AOT and AI-DUST predicted dust concentration over North Africa and the Arabian Peninsula on 22 March and 12 May 2025.

figure 11

a, c show VIIRS Deep Blue AOT observations; b, d show AI-DUST forecasts initialized at 00:00 UTC on 21 March and 11 May, respectively. Model forecasts are sampled at times closest to satellite overpass for direct comparison.

On 22 March, a prominent dust plume extended from the central Sahara (0°–20°E) northward, as evidenced by high AOT in satellite observations. Despite data gaps due to partial cloud cover, the spatial structure of this northward-traveling SDS band was clearly identifiable. Additionally, high AOT was observed near 30°E in the southern Sahara. AI-DUST successfully reproduced the SDS feature with strong spatial agreement. However, it slightly overestimated SDS in the 30°E region, indicating possible limitations in local emission parameterizations.

On 12 May, intense SDS activity occurred across two major zones, the northern and western Sahara, and the northeastern Arabian Peninsula. The AI-DUST model accurately reproduced the location and spatial extent of both SDSs, demonstrating strong predictive skill despite the geographic and meteorological differences from its training domain.

Overall, the model’s performance in these subtropical and tropical arid regions is highly encouraging. Its ability to reproduce SDS features without retraining or regional tuning demonstrates robust generalization and geographical portability. This cross-regional success suggests that AI-DUST has learned fundamental physical patterns governing dust diffusion, transport, and deposition, rather than merely memorizing region-specific statistics, highlighting its potential as a globally applicable tool for SDS forecasting.