Introduction

Recent advancements in Artificial Intelligence (AI) global weather forecasting have marked a major breakthrough in the realm of weather forecasting, ushering in capabilities that surpass traditional global forecast models1. AI weather forecasting, leveraging sophisticated algorithms and extensive data analysis, has evolved rapidly, offering enhanced accuracy and reliability in predicting various weather phenomena2,3,4. The ability of AI weather forecasting to exceed the performance of established global models signifies a pivotal shift in meteorological methodologies, opening new avenues for research and application in this field5,6. Recognizing its potential to markedly enhance forecast precision, meteorological centers like the European Centre for Medium-Range Weather Forecasts (EC) and the China Meteorological Administration (CMA) are actively exploring and integrating AI into their global weather forecasting systems7.

Despite these considerable strides, AI weather forecasting still faces challenges, such as in accurately predicting typhoon intensity, which is notorious for its complex structures and rapid intensity changes8,9. AI models also encounter limitations in resolution and struggle with the nuances of regression algorithms, leading to an underestimation of typhoon intensity and insufficient representation of storm structures6,10. Consequently, a critical gap exists in AI’s ability to model the intricacies of typhoon dynamics comprehensively.

Similarly, global numerical weather prediction (NWP) models have historically also encountered a considerable challenge in accurately forecasting typhoon intensity due to resolution constraints11,12. For the typhoon intensity forecast, a well-established and widely accepted approach to address this issue primarily relies on the use of regional NWP models, complemented by the integration of various vortex initialization schemes13,14,15. The integration of vortex initialization into these regional models has markedly enhanced their capabilities in forecasting typhoon intensity, leading to their widespread adoption in advanced operational systems in the United States12 and in China16,17,18. This naturally raises an intriguing question: could a novel strategy be applied to enhance AI weather forecasting for typhoon intensity prediction? Considering that both AI and traditional global NWP models face challenges related to resolution constraints, the incorporation of vortex initialization within regional models has the potential to improve the accuracy of AI models in predicting typhoon intensity. This consideration at least opens up the possibility of leveraging the strengths of both AI and traditional numerical modeling techniques to develop a more robust and accurate system for predicting the typhoons intensity.

This study proposes an alternative approach in which Pangu-weather (hereafter, Pangu)6, a state-of-the-art AI weather forecasting system, is deployed to provide initial fields and lateral boundary conditions to drive the regional NWP, such as Weather Research and Forecasting (WRF) model. This innovative framework, designated as ‘AI-Driven WRF’, stands in contrast to the conventional approach powered by global numerical models (‘NWP-Driven WRF’). By integrating the high computational efficiency and precise larger-scale forecasting of advanced Pangu19,20 with the finer-scale prediction capabilities of WRF models21,22,23,24 and dynamic vortex initialization (DI), the study overcomes AI’s limitations in accurately capturing the full complexity of typhoon dynamic systems. This research endeavors to enhance the precision of typhoon intensity predictions and contributes to the broader field of meteorology by demonstrating the potential of AI forecasting models in comprehending and forecasting typhoon intensity.

Results

Comparative analysis of forecast models and dynamic initialization effects

The comparative assessment delineates salient contrasts between the Pangu AI model’s global forecasts and the regionally focused, AI-driven WRF simulations, elucidating their respective implications for atmospheric predictive accuracy. The Pangu model is well-suited for the delineation of synoptic-scale meteorological patterns as the flowchart depicted in Fig. 1a. Nonetheless, the limited horizontal resolution makes it hard to resolve mesoscale processes, particularly for convective and extreme weather systems, as evidenced by the underestimated typhoon vortex intensity (Fig. 1d).

Fig. 1: Comparative analysis of forecast models and initialization effects.
figure 1

a The flowchart of the Pangu Global Weather Forecast Model as developed by Bi et al.6, b defined domain of the AI-driven WRF Model, c the flowchart of the vortex initialization (DI) scheme, d vortex structure from Pangu, and e vortex structure after DI depicted by the wind speed at the lowest model level recorded of typhoon Doksuri at 0000 UTC on 24 July 2023.

Full size image

In contrast, the WRF model’s fine-resolution could extend to 2 km through its multi-nesting capabilities (Fig. 1b), provides an intricate depiction of local-scale weather phenomena. This high resolution is imperative for the veracious analysis and severe weather events forecast, including typhoons25. Moreover, applying dynamic initialization (Fig. 1c) in WRF model could further refine the AI-forecasted vortex in the initial conditions (Fig. 1d), enhancing the intensity and structure of vortex formations to yield a more realistic and reliable prediction (Fig. 1e).

Therefore, while Pangu model provides an essential foundation for global weather forecasting, WRF model, with its superior resolution and dynamic initialization, offers an invaluable tool for improving typhoon intensity simulation as Pangu AI global model forecasted. This enhancement holds considerable promise in markedly refining Pangu’s utility in providing initial conditions for WRF model.

Comparative validation of the Pangu and EC models as initial fields

The actual application of the Pangu model as the WRF’s initial field is based on validating its initial accuracy, especially against the current mainstream technique of the global NWP model. EC, renowned global forecasting performance26, is selected as the benchmark to compare with Pangu model. Together with ERA5 reanalysis data, the techniques of Pangu and EC models can accurately simulate the column-integrated moisture flux and its corresponding magnitude at 0000 UTC on 24 July 2023 in case of Typhoons Doksuri and Khannu (Fig. 2a–c). Notably, Khannu was a concurrent typhoon that may have interacted with Doksuri, presenting a unique scenario for analyzing the models’ capabilities in simulating complex atmospheric interactions between coexisting typhoons. Furthermore, the analysis also highlights both techniques effectiveness in capturing monsoonal moisture transport and accurately positioning substantial weather systems, such as Typhoons Doksuri and Khannu near the subtropical high. In addition, the differences between the two techniques and ERA5 featured by moisture flux and its magnitude further illustrate that both the Pangu and traditional NWP models overall reflect the observation (Fig. 2d, e). To further extend the effectiveness of both techniques, the same performance is applied to a supplementary case of Typhoon Hato (see Fig. S1), still exhibiting close correspondence with the ERA5 datasets. Therefore, the result underscores AI model’s accuracy in forecasting large-scale weather patterns, affirming its utility as initial fields. However, both techniques obviously underestimate the typhoon intensity as shown in Fig. 2d, e. Considering that ERA5 data inherently underestimates typhoon intensity due to its resolution, this suggests that the intensity forecasts by both Pangu and EC models are noticeably underestimated. Therefore, to better leverage the larger-scale information provided by these models when using them as initial fields for WRF-based typhoon intensity predictions, it becomes imperative to employ the vortex initialization technique to further enhance forecasting accuracy (Fig. S2).

Fig. 2: Comparative validation of the Pangu and EC models as initial fields.
figure 2

Column-integrated moisture flux (vectors) and its corresponding magnitude (shading; units: kg m−1 s−1) for the a ERA5, b 24-h forecast by the Pangu model, and c 24-h forecast by EC at 0000 UTC on 24 July 2023. The difference in the column-integrated moisture flux and its magnitude at 0000 UTC on 24 July 2023 between 24-h forecasts from d the Pangu model, and e EC and the ERA5 reanalysis dataset. Vertical profiles of the RMSE for f geopotential height (units: m2 s−2), g meridional wind (units: m s−1), and h zonal wind (units: m s−1) associated with Typhoon Doksuri (represented in blue) and Typhoon Hato (depicted in green), within the range of 0–40°N and 90–150°E. The labels ‘Doksuri’, ‘Khannu’, ‘Monsoon’, and ‘SH’ in Fig. 2a indicate Typhoons Doksuri and Khannu, monsoonal moisture flux, and the subtropical high, respectively.

Full size image

In addition, we also conducted further validation of the Pangu and EC models against radiosonde observations to evaluate their proficiency in simulating Typhoons Doksuri and Hato (Fig. 2f, g). Results show that the Pangu model consistently outperformed the EC model, demonstrating enhanced atmospheric simulation capabilities attributed to better simulating atmospheric dynamics. Compared to the EC model, the RMSE of the Pangu model reduces up to 2 m² s−2 for geopotential height and 0.5 m s−1 for both meridional and zonal wind fields. Furthermore, consistent patterns in the vertical profiles of Typhoons Doksuri and Hato corroborated the Pangu model’s accuracy, emphasizing its effective representation of large-scale circulations during typhoon conditions. These findings hint the Pangu model’s potential as a robust and reliable foundation for initializing the WRF regional model, enhancing our capacity for accurate typhoon forecasting.

Comparative analysis of typhoon track and intensity

Following the establishment of the Pangu model as an appropriate initial field for the WRF model, we conducted numerical simulations to assess and compare the impact of the AI-Driven WRF approach on typhoon intensity and structure of Typhoons Doksuri and Hato. These simulations, utilizing a triple-moving-nested grid, implemented the same physics package as described in the method part. Hence, this section’s main point is to assess the WRF model’s forecasting proficiency with presenting a detailed comparative assessment of the intensity evolution along the track simulation of Typhoons Doksuri and Hato (Fig. 3).

Fig. 3: Comparative analysis of typhoon track and intensity.
figure 3

a, b The track simulation of typhoons Doksuri (a) and Hato (b) against the best track. c, d The evolution of the maximum 10-m wind of typhoons Doksuri (c) and Hato (d).

Full size image

In the case of Typhoon Doksuri, initial simulations up to 24 h closely mirror the typhoon’s actual track (Fig. 3a). However, discrepancies become evident after 24-h, with the EC_WRF simulation suggesting a westward turn while the Pangu and Pangu_WRF simulation veers more to the east. Given that the Pangu_WRF model is driven by the Pangu forecasts, their predicted paths exhibit a notable correlation, yet the application of vortex initialization at the onset enhances the track precision over the original Pangu predictions. Regarding maximum surface wind evolution, both EC_WRF and Pangu_WRF runs show rapid intensification, but the Pangu_WRF aligns more closely with the best track record, with the RMSE of intensity reducing from 6.1 to 5.1 m s−1, a marked improvement over Pangu’s RMSE of 29.3 m s−1 (Fig. 3c).

For Typhoon Hato, the Pangu_WRF and Pangu simulation offers a more accurate track reproduction for most of the simulation period, staying closer to the Joint Typhoon Warning Center compared to the EC_WRF simulation (Fig. 3b). This improved track simulation with Pangu_WRF is also accompanied by a more accurate forecast of typhoon intensity. The maximum surface winds simulated by Pangu_WRF were closer to the best track data, with a reduced RMSE from 10.6 to 6.1 m s−1, in contrast to Pangu’s RMSE of 16.1 m s−1 (Fig. 3d). Note that, results reveal significant fluctuations in the simulated tropical cyclone intensities. These fluctuations in Pangu_WRF and EC_WRF, are likely due to our method of calculating typhoon intensity, which involves recording at 15-min intervals and applying a four-point smoothing process. This method inherently introduces fluctuations in both models, with Pangu_WRF displaying greater variability due to its stronger intensity.

The advancements in typhoon intensity simulation facilitated by Pangu_WRF, compared to the original Pangu model, are significant, with intensity forecasts for Typhoons Doksuri and Hato seeing improvements of 87% and 63%, respectively. Alongside this substantial progress, there’s a considerable reduction in track error in early simulations. This improvement can be attributed to two factors. One, the significant enhancement in resolution provided by WRF allows for the resolution of finer-scale structures, leading to more detailed and accurate simulations. The other, the dynamical initialization improves the initial position of the typhoon, resulting in more accurate forecasts. These results highlight the potential of AI-Driven WRF to enhance the accuracy of both typhoon intensity and track forecasting.

Assessing typhoon intensity with SAR data

In this section, we shifted our focus to the examination of the WRF model’s effectiveness in predicting the intensity of typhoon with Synthetic Aperture Radar (SAR) data. The presence of detailed high-resolution SAR wind datasets, especially for Typhoon Doksuri in 2023, presented a unique chance to corroborate the structures of typhoons as simulated by the WRF model27. Nevertheless, it is important to note the absence of SAR wind data for Typhoon Hato, limiting our capability to perform a comparable analysis for this typhoon.

Figure 4 presents a detailed comparison of horizontal wind speed distributions at 10-m, exhibiting SAR results (Fig. 4a, d) against simulations from the Pangu_WRF (Fig. 4b, e) and EC_WRF (Fig. 4c, f) during critical phases of Typhoon Doksuri at 1000 UTC on 24 July and 25 July 2023. The SAR data highlights intense wind speeds near the surface, serving as a benchmark for evaluating the simulations. Transitioning from the Pangu model to the Pangu_WRF simulation reveals a significant uptick in the accuracy of capturing Typhoon Doksuri’s intensity, with Pangu_WRF showing a marked improvement in simulating the typhoon’s dynamics more faithfully than its Pangu. For example, on 24 July, SAR data recorded intensity in terms of maximum wind speeds of 48.3 m s−1, with the initial Pangu model and Pangu_WRF simulations producing wind speeds of 23.9 m s−1 and 49.0 m s−1, respectively, while EC_WRF recorded 44.3 m s−1.

Fig. 4: Typhoon Doksuri structure validation analysis.
figure 4

The horizontal distributions of wind speed (m s−1) at an altitude of 10 m, recorded at 1000 UTC on 24 July 2023, across a SAR, b Pangu, c Pangu_WRF, and d EC_WRF simulations. e–h The same parameters and datasets as (a–d) but recorded for 1000 UTC on 25 July 2023.

Full size image

On 25 July, the intensity increased with SAR documenting 57.2 m s−1; the Pangu model registered 26.2 m s−1, Pangu_WRF closely matched the observed intensity with 59.1 m s−1, and EC_WRF showed a slight underestimation with 56.9 m s−1. Although the maximum wind speed simulated by EC_WRF was closer to the SAR observations, the overall wind intensity within the eyewall, as simulated by Pangu_WRF, more accurately reflected the observed conditions as shown in the last two columns of Fig. 4 indicating a more comprehensive match to the SAR data (Fig. 4e, h). While the inner-core structure of EC_WRF (Fig. 4h) seems to match SAR observations (Fig. 4e) more closely, the strong wind field exceeding 50 m s−1 and the asymmetrical wind structure around 35 m s−1 are better represented by Pangu_WRF (Fig. 4g). These results in figures, on average across the two evaluated moments, underscore the potential of Pangu_WRF in typhoon forecasting, not only surpassing the original Pangu AI model but also offering potential improvements over EC_WRF.

While Pangu_WRF and EC_WRF have made strides in aligning more closely with SAR-derived wind speeds, they both encounter the challenge of eyewall size overestimation in comparison to SAR data. This issue, as investigated by Xu and Wang28 and Xu and Duan29, suggests that adopting finer, nearly sub-kilometer scale resolutions could significantly enhance the depiction of the eyewall’s complex structure. This indicates that enhancing the resolution of regional models could further elevate the accuracy of AI-based typhoon intensity forecasts.

Environmental accuracy and its impact on typhoon intensification

Although the AI model generates smoother outputs compared to traditional numerical global models, it still exhibits significant potential in intensity forecasting—how is this achieved? Quantitative evaluations presented in Fig. 2 demonstrate that the Pangu model provides a more accurate depiction of the environmental field compared to the EC model, particularly in terms of wind field, moisture transport, and geopotential height around Typhoon Doksuri. For instance, the Pangu model’s outputs show better alignment with the ERA5 reanalysis data (Fig. 5a–c), notably in moisture transport near 10°N and the geopotential height field on the typhoon’s east side. These results underscore that, despite the relatively smoother and coarser resolution compared to global NWP models, the Pangu-weather model’s enhanced capability in simulating larger-scale circulation (e.g., >300 km) compensates for its limitations20, which is critical for typhoon development.

Fig. 5: Comparison of environmental accuracy and typhoon vertical structure.
figure 5

a–c Moisture flux magnitude (shading; units: kg kg−1 m s−1) and geopotential height (gpm) at 700 hPa for a ERA5, b Pangu, and c EC, at 0000 UTC on 24 July 2023. d, e Azimuthally averaged radius–height cross-sections tangential wind speed (m s−1, black contours1) and radial wind speed (m s−1, shaded) recorded at 0000 UTC on 24 July 2023 for the d Pangu_WRF and e EC_WRF experiments.

Full size image

This realistic simulation of environmental fields, in turn, enables more effective dynamical initialization, resulting in a more favorable initial vortex structure for intensification. As shown in Fig. S3, while the maximum wind speeds of Pangu_WRF and EC_WRF are similar, their horizontal and vertical structures exhibit substantial differences. The Pangu_WRF model features a more compact eyewall structure, stronger outflow at the top of the troposphere, and more developed tangential winds extending to higher altitudes (Fig. 5d, e)—all of which are conducive to typhoon intensification.

Therefore, during the vortex initialization phase, despite initial discrepancies in typhoon intensity, the accurate environmental forecasts from the Pangu model facilitate the development of a more favorable typhoon vortex. In the subsequent forecast phase, this improved vortex structure, combined with precise large-scale information from Pangu, promotes more rapid intensification. This mechanism explains why the Pangu_WRF model’s intensity forecasts were superior to those of the EC_WRF model in current study.

Validation of typhoon cycle forecasts

The above analysis demonstrates that using the Pangu AI model as a driving field for the WRF model is both feasible and effective. However, a key question remains: can the Pangu-driven WRF also perform effectively in operational scenarios requiring batch cycle runs? To address this, we conducted a series of batch tests on two typhoons from 2023, Doksuri and Khanun, using a similar setup as described in the methods section. To better replicate operational conditions, we initialized the vortex at the analysis time and started the forecast 6 h later, incorporating the vortex into the 6-h forecasts. In this study, we compared EC forecasts with those generated by Pangu, with both models producing forecasts, each covering a 120-h period. Due to limitations in obtaining EC forecast data from the CMA, forecasts could only be initiated twice daily.

Our results reveal that the Pangu_WRF consistently performed slightly better than EC_WRF in terms of intensity forecasting across these batch cycles (Fig. 6a, b). For instance, in the Doksuri experiment, the average intensity error for the 0–72 h forecast decreased from 6.6 m s−1 with EC_WRF to 6.2 m s−1 with Pangu_WRF (Fig. 6c). For Typhoon Khanun, the intensity error decreased from 9.6 to 9.4 m s−1, suggesting a slight improvement in intensity forecasting accuracy with Pangu_WRF (Fig. 6d).

Fig. 6: Validation of typhoon cycle forecasts.
figure 6

a, b Temporal evolution of maximum surface wind speed for Typhoons Doksuri and Khanun, compared with JTWC best-track data. c, d Intensity, and e, f position errors for Typhoons Doksuri and Khanun, respectively.

Full size image

Regarding track error, for Typhoon Doksuri, the 0–72-h average track error was reduced from 76 km with EC_WRF to approximately 60 km with Pangu_WRF (Fig. 6e). However, for Typhoon Khanun, the average track error increased from 97.1 to 106.4 km (Fig. 6f). This increase in track error for Khanun may be attributed to a slightly larger initial vortex position error in Pangu, likely due to its lower resolution, which compounded the subsequent errors. Despite this, for shorter-term forecasts (0–48 h), the differences between the two models were minimal.

To assess whether these results persist over longer lead times, we extended the forecasts to 120 h and analyzed their intensity errors. Forecast performance can vary noticeably among different typhoons—a phenomenon that is understandable since small initial differences can amplify over longer forecast intervals. For Doksuri, the intensity error in Pangu_WRF at 120 h was 6.3 m s−1, compared to EC_WRF’s 9.1 m s−1, representing a 30.8% reduction in error. In contrast, for Khanun, the intensity error at 120 h was 10.7 m s−1 for Pangu_WRF versus 9.9 m s−1 for EC_WRF, showing an 8.1% increase. These extended forecasts underscore the inherent variability of AI-driven simulations at longer lead times, yet they also highlight the AI-driven approach’s overall comparability and its potential when measured against traditional NWP-driven methods.

Despite these variations between different typhoons, the overall performance of the Pangu-driven WRF model is comparable to that of the EC-driven WRF. These findings further demonstrate the feasibility and effectiveness of the Pangu-driven WRF approach for forecasting typhoon intensity, suggesting that it could be a viable alternative in operational meteorology.

Discussion

This study addresses the substantial challenges inherent in enhancing AI weather forecasting for accurate typhoon intensity prediction. The limitations of current models, chiefly due to their algorithmic constraints and relatively coarse resolution, impede their effectiveness in this field. Our research introduces a novel approach by applying AI weather forecasting to drive regional WRF model, further augmented by the incorporation of DI technique. This methodology substantially elevates the proficiency of AI weather forecasting in precisely predicting the intensity of typhoons, an advancement of paramount importance considering the potentially catastrophic impact of these natural events.

Our evaluation involved three distinct experiments: The initial experiment, using the Pangu AI weather forecasting model, revealed their limitations in accurately predicting typhoon intensity due to the models’ coarse resolution, leading to a significant underestimation of the typhoons’ intensity and poor distinguishing wind structural details. The latter two experiments utilized the traditional NWP global model (EC-Driven) and the Pangu AI global model (AI-Driven) WRF model, both augmented by dynamic initialization. Both EC-driven and AI-Driven WRF models demonstrated exceptional capability in capturing the complex details and dynamics of typhoons, yielding more precise predictions of intensity. Additionally, when compared with the EC-driven WRF model, AI-Driven WRF model showed potential improvements in forecasting accuracy in terms of intensity. This highlights the substantial potential of AI-driven models in improving weather predictions, especially for severe weather events like typhoons.

As an exploratory study, our primary goal was to demonstrate the feasibility and potential of the AI-driven WRF approach, which represents a significant advancement in forecasting typhoon intensity. Our preliminary experiments with Typhoons Doksuri (2023) and Khanun (2023) indicate that this approach holds considerable promise. However, while typhoon intensity forecasting involves multi-scale interactions, the AI weather forecasting model exhibits significantly higher accuracy in predicting larger-scale processes but lower capability in resolving finer scales compared to traditional global numerical forecasting models22. Therefore, determining whether the forecasts are improved or significantly enhanced will require a more extensive set of test cases. Future research should prioritize a comprehensive evaluation of the model across a broader range of scenarios and focus on refining the algorithms, particularly to address the complex dynamics of typhoon systems. This will involve rigorous testing under diverse and challenging weather conditions, as well as exploring further enhancements, such as incorporating higher-resolution AI models. The ongoing development and evaluation of this approach are expected to contribute substantially to the field of meteorology, enabling more reliable and actionable forecasts for typhoon intensity and improving our capacity to mitigate the impacts of such events effectively.

Methods

Datasets

The EC ERA5 reanalysis data, known for its comprehensive assimilation of observational data and high accuracy26, was employed to drive the Pangu model and assess its forecasting performance. In Fig. 2a–e, ERA5 reanalysis data was used to evaluate the forecast differences between the AI and EC models. However, recognizing that ERA5 is not direct observational data, we used the NCEP Global Upper Air Observations in Fig. 2f–h for a further quantitative evaluation of the forecast differences between the AI and EC models. Due to the sparse radiosonde coverage over oceanic regions (Fig. S5), it remains challenging to fully capture the typhoon environment. By combining these radiosonde observations with ERA5 reanalysis, we added an additional layer of validation to ensure the robustness of our comparisons. Additionally, Synthetic Aperture Radar (SAR) wind data, valued for its high-resolution and precise wind pattern depiction, was crucial for verifying the structural characteristics of typhoons. The study also utilized the Joint Typhoon Warning Center Best Track Archive to validate the model’s predictions regarding typhoon intensity and path, thus offering a solid foundation for appraising the model’s performance.

Pangu-weather forecasting

Pangu represents a revolutionary development in AI global weather forecasting, crafted by Bi et al.6 (Fig. 1a). It’s a deep learning-based system trained extensively on ERA5 reanalysis data from 1979 to 2017, encompassing over 341,880 hourly data points. Pangu’s design includes four deep neural networks, each specialized for different forecast lead times: 1, 3, 6, and 24 h. In this study, to maintain the continuity of the results, we utilized 6-h intervals to drive the WRF model. Its core innovation lies in the 3D Earth-specific transformer, enabling the effective integration of various meteorological variables into a cohesive network. This approach involves down-sampling data into a 3D cube format, processed via an encoder-decoder structure. The unique Earth-specific positional bias introduced by this transformer enhances forecast accuracy while maintaining computational efficiency. Notably, Pangu has demonstrated superior performance over the globally esteemed ECMWF operational system, particularly in deterministic forecasts and extreme weather predictions, including tropical cyclone tracking. However, a notable limitation of Pangu lies in its resolution with 0.25° × 0.25°.

It is important to note that while Pangu-weather model provides all necessary upper-air variables (u, v, t, z, q) to drive the WRF model, it lacks certain surface variables such as sea surface temperature (SST), soil moisture, and soil temperature. SST is particularly crucial for accurate typhoon intensity forecasts; thus, in operational settings, high-resolution SST data is often employed to ensure precision. In our experiments, we used ERA5 reanalysis data to set these surface variables, with the SST remaining fixed throughout the simulation. Consequently, the current set of variables from Pangu, when supplemented with data from other NWP models, is sufficient for operational use. Incorporating these additional variables into Pangu is technically feasible, and future collaboration with AI developers can further enhance the integration of AI-generated data into regional numerical models.

WRF model

The Advanced Research WRF model (version 4.5.2) was driven by the Pangu and the operational integrated forecasting system of EC for a detailed analysis of typhoon intensity and structure. The model was set up with three domains (D01–D03), each featuring distinct horizontal grid sizes and grid spacing as exhibited in Fig. 1b: D01 with 311 × 251 grid points (18 km), D02 with 271 × 271 (6 km), and D03 with 211 × 211 (2 km). Notably, D02 and D03 were specifically designed as vortex-following moving grids, ensuring comprehensive coverage and a refined representation of the typhoon’s circulation. The vertical stratification of the model comprised 50 levels, extending up to a top boundary of 50 hPa.

For the model physics process, the WRF-single-moment-microphysics class 6 (WSM6) scheme30 was chosen. Radiation parameterization was conducted using Dudhia shortwave31 and Rapid Radiative Transfer Model (RRTM) longwave32 schemes. Cumulus parameterization was only applied to the outermost domain (D01) using the Kain–Fritsch scheme33,34. The surface layer was parameterized employing the revised MM5 Monin-Obukhov scheme. Considering the pivotal importance of model physics in precisely depicting typhoon intensity and structure, all experiments in this study uniformly utilized these specified schemes.

Note that Pangu model has only 13 vertical levels, which may affect the model results to some extent. However, when using global forecast data to drive a regional model, it is common for the global model to have fewer vertical levels than the regional model. Interpolation is necessary to align the initial conditions. For WRF, we used its real.exe program to interpolate the pressure levels from both the EC and Pangu-weather global models to WRF’s vertical coordinate levels. This interpolation process ensures that the vertical structure is adequately initiated. Future improvements could include using AI models with more vertical levels, such as Fengwu and Graphcast.

Dynamical initialization scheme for typhoons

To enhance the initial vortex structure in the model, this study implemented the dynamic vortex initialization (DI) scheme developed by Cha and Wang13 for the WRF Model (Fig. 1c), specifically aimed at typhoon intensity forecasting. This comprehensive scheme involves three key steps: firstly, Separate typhoon vortex: vortex separation using local smoothing and azimuthal averaging techniques to isolate the typhoon’s axisymmetric component; secondly, Spectral nudging: dynamically enhancing this component through a series of 6-h integrations within the WRF Model, starting 6 h ahead of the forecast, with the application dependent on the typhoon’s intensity; and thirdly, Relocate spun-up typhoon to observed position: a crucial relocation procedure to align the typhoon with its observed geographical position, ensuring accurate initial conditions. Building on this approach, Liu et al.11 further refined the DI scheme for application in complex terrain, enhancing its overall effectiveness. This advanced implementation of the DI scheme substantially improves the creation of realistic initial conditions for typhoons and reduces forecasting errors, finally improving the accuracy of forecasts within the WRF framework.

Experimental design

Numerical simulations were implemented to evaluate the AI-Driven WRF model’s performance, targeting two substantial typhoons: Doksuri, from 0000 UTC 24 to 0000 UTC 27 July 2023, and Hato, from 1800 UTC 21 to 1800 UTC 24 August 2017. Typhoons Doksuri and Hato were selected for their substantial impacts35,36 and the alignment of their development stages with the initialization times of the regional model, thereby providing a robust test for the AI-driven WRF model’s capability to predict typhoon intensity.

The Pangu experiment employed ERA5 reanalysis data as the initial condition for the model’s forecasting. In addition, two distinct experiments, EC_WRF and Pangu_WRF, were conducted. The EC_WRF experiment applied traditional global NWP technique, utilizing global forecasting data from the EC integrated forecasting system for its initial and lateral boundary conditions (EC-Driven WRF), serving as a representative instance of NWP-Driven WRF. In contrast, the Pangu_WRF experiment employed the data from Pangu-weather AI forecasting model for setting its initial and lateral boundary conditions (Pangu-Driven WRF), illustrating an application of AI-Driven WRF. Given that Pangu model uses ERA5 data as its driving initial field, 24-h forecasts from both EC and Pangu with 6-h intervals were used to explore differences in initial conditions. This choice was made to highlight the differences and effectiveness of using AI-generated data versus traditional NWP data for driving the WRF model. Both Pangu and EC models were initiated at 0000 UTC 23 July 2023 for Doksuri and 1800 UTC 20 August 2017 for Hato. The EC_WRF and Pangu_WRF experiments were then initiated at 0000 UTC 24 July 2023 for Doksuri and 1800 UTC 21 August 2017 for Hato, respectively. To ensure a fair comparison between the two experiments, both employed dynamic initialization and used the same model settings. The key difference between these two experiments is solely in their initial and lateral boundary conditions. This consistent setup facilitates a focused assessment and comparison of the effectiveness of the conventional EC- and Pangu-Driven weather forecasting approaches, particularly in their accuracy in forecasting the typhoon intensity.