Over decades, numerical weather prediction has been dominated by physics-driven regional models. These models rely on solving complex equations that describe atmospheric dynamics and thermodynamics. While sophisticated parameterization schemes and ensemble forecasting have enhanced their performance, the computational demands have ballooned, challenging real-time operational feasibility. Increasing model resolution and extending ensemble sizes have inevitably led to higher costs, limiting the frequency and scope of accurate forecasts. This backdrop sets the stage for integrating advanced artificial intelligence techniques into meteorology.

The infusion of machine learning, particularly deep learning architectures like convolutional neural networks (CNNs) and transformers, into weather prediction illustrates a paradigm shift. Early machine-learning weather prediction models proved the concept that data-driven methods could analyze vast meteorological datasets and generate rapidly executed forecasts. However, these models excel primarily at coarse, large-scale atmospheric features and often fall short in resolving mesoscale or convective-scale phenomena essential for accurate typhoon intensity and structure predictions. The challenge remains to balance computational speed with physical fidelity in forecasting models.

The Shanghai team’s bold question was whether a hybrid approach, blending the complementary strengths of physics-based and AI-driven models, could overcome these limitations. Physics models effectively simulate mesoscale typhoon structures, while artificial intelligence models better capture voluminous large-scale atmospheric flows. Combining these two fronts could achieve a synergistic enhancement in forecast accuracy, particularly for potentially catastrophic events like typhoons.

Machine learning models developed worldwide, including PanGu, GraphCast, FengWu, FuXi, and the Artificial Intelligence Forecasting System, largely employ the transformer architecture, known for handling sequential data and long-range dependencies. While outperforming traditional physics models on large-scale metrics, they struggle with mesoscale details and often underestimate tropical cyclone intensities. Despite their lightning-fast forecast generation, the high training costs and ongoing accuracy challenges underscore the need for innovative hybrid methodologies.

The Shanghai Typhoon Model embodies this innovation by embedding machine-learning components within a physics-driven framework, creating a hybrid system that leverages the physics model’s mesoscale resolving power and the AI’s ability to capture large-scale atmospheric dynamics. This fusion empowers the model to maintain mean track forecast errors below 200 kilometers for extended lead times up to 108 hours, a feat that surpasses even ECMWF’s Integrated Forecasting System and state-of-the-art ML models like PanGu and FuXi during Typhoon Danas.

Typhoon Danas, which occurred in mid-2025, offered an exemplary testbed for this hybrid forecasting system. Using sequential forecasts initialized from July 5 to July 7, captured vividly in FY-4B satellite visible light imagery, the hybrid model consistently delivered superior track predictions compared to the purely physics-based or fully machine-learning approaches. This robustness in sustained accuracy indicates the hybrid model’s operational viability for real-time typhoon forecasting.

Wei Huang, a leading researcher at the Shanghai Typhoon Institute, emphasized that the hybrid model “used artificial intelligence to anchor the large-scale flow and physics to resolve mesoscale structure,” tackling the core challenge of mesoscale prediction skill that pure ML models face. The success with Danas points toward a new operational forecasting paradigm where hybrid models provide both accuracy and computational efficiency, essential for disaster preparedness and response.

Despite remarkable progress, the hybrid system is not without limitations. Early lead-time forecasts still encounter some challenges, and computational efficiency, while improved, remains constrained by the underlying physics-based elements. Addressing these issues inspires the team’s forward-looking ambition—to develop a purely data-driven, regional machine-learning typhoon forecasting model that maintains uncompromised physics credibility and dramatically enhances forecast speed without sacrificing accuracy.

To this end, ongoing research focuses on refining initial conditions, improving data assimilation techniques, and exploring scenario-specific machine-learning models tailored for tropical cyclone dynamics. This trajectory aligns with the broader meteorological community’s shift toward AI-augmented weather prediction frameworks, anticipating an era where machine learning fully complements or even replaces traditional physics-based models in operational forecasting.

The implications extend beyond regional forecasting: advances cultivated in the Shanghai Typhoon Model set a precedent for hybrid and pure data-driven approaches in other severe weather phenomena worldwide. By marrying the deep physical understanding of the atmosphere with the pattern recognition and computational efficiency of AI, the next generation of weather models could revolutionize early warning systems, risk management, and climate resilience strategies on a global scale.

The team’s full findings were published in the September 18, 2025, issue of Advances in Atmospheric Sciences, marking a milestone in tropical cyclone research. Their work underscores not only the scientific advancement but also the necessity of interdisciplinary collaboration between meteorologists, computer scientists, and data engineers to conceptualize and implement these complex, hybrid forecasting systems.

Looking ahead, the Shanghai team remains committed to pushing the boundary from hybrid towards fully data-driven typhoon forecasting models, while ensuring physical realism remains integral. Success in this endeavor will hinge on advances in both physical modeling—enabling more credible training datasets—and machine learning methodology, ultimately enhancing society’s preparedness against the increasing threats posed by tropical cyclones amid climate change.

As weather prediction enters this new frontier, the Shanghai Typhoon Model’s evolution epitomizes the transformative potential of artificial intelligence harmonized with established physical science. Their groundbreaking case study of Typhoon Danas stands as a beacon illuminating the path toward the future of high-accuracy, efficient, and reliable typhoon forecasting.

Subject of Research: Development and evaluation of hybrid and machine-learning typhoon forecasting models, with a case study on Typhoon Danas (2025).

Article Title: Evaluating the Shanghai Typhoon Model against State-of-the-Art Machine-Learning Weather Prediction Models: A Case Study for Typhoon Danas (2025)

News Publication Date: September 18, 2025

Image Credits: Northwest Pacific Tropical Cyclone Retrieval System, Shanghai Typhoon Institute

The core advancement realized with AIFS ENS lies in its ensemble approach to AI-powered weather modeling. Unlike deterministic forecasts, which produce a single projected atmospheric outcome, ensemble forecasting generates multiple simulations with slight perturbations in initial conditions. This technique captures the inherent uncertainty of weather systems, providing meteorologists and stakeholders with probabilistic insights that are crucial for informed decision-making. ECMWF’s AIFS ENS is a milestone because it successfully integrates AI and machine learning technologies within the ensemble forecasting framework, a method ECMWF has pioneered and refined over the last three decades.

From a technical standpoint, the AIFS ENS leverages the immense data assimilation capabilities characteristic of physics-based models to establish accurate initial atmospheric states. By using these rigorous physics-driven initializations as inputs, the AI model then executes rapid forecast simulations that are computationally efficient and energy-conscious. ECMWF reports that AIFS ENS achieves forecast generation over ten times faster than traditional ensemble methodologies, while reducing computational energy consumption by a factor of approximately one thousand. This breakthrough brings significant benefits not only in forecast timeliness but also in sustainability, addressing the ever-growing ecological footprint of large-scale numerical weather prediction operations.

Despite these impressive gains, ECMWF recognizes that the AI-driven ensemble model currently operates at a spatial resolution of approximately 31 kilometers, which remains somewhat coarser compared to their state-of-the-art physics-based ensemble systems. The latter remains unmatched for high-resolution weather parameterizations and coupled Earth system modeling, which are indispensable for capturing finely detailed atmospheric phenomena and interactions between the atmosphere, ocean, and land surfaces. Therefore, ECMWF is actively exploring hybrid forecasting paradigms that synergize AI’s speed and accuracy with the granular physical fidelity of traditional models.

The innovation embedded in AIFS ENS is aligned with ECMWF’s larger vision of harnessing machine learning to transform meteorological forecasting. Earlier in the year, ECMWF pioneered the first operational data-driven forecasting model, AIFS Single, which executes single forecast runs rapidly and accurately but lacks the probabilistic nuance of ensembles. The ensemble expansion with AIFS ENS therefore addresses a critical demand from meteorological services and users who require comprehensive scenario analysis rather than deterministic projections, improving risk assessment in sectors ranging from agriculture and energy to disaster preparedness.

ECMWF’s Director-General, Dr. Florence Rabier, highlighted the collaborative and scientific significance of this achievement. Dr. Rabier emphasized that the operationalization of a 51-member ensemble AI forecasting system is a landmark for ECMWF and its Member States. The accessibility of AIFS ENS as an open-source tool exemplifies ECMWF’s commitment to international cooperation among its 35 Member and Co-operating States, empowering national weather services to enhance prediction accuracy and public safety worldwide. This democratization of advanced AI forecasting infrastructure is poised to provide a transformative impact on global weather preparedness.

Echoing this vision, Dr. Andy Brown, ECMWF’s Director of Research, underscored the scientific rigor behind AIFS ENS, noting the model as emblematic of ECMWF’s dedication to innovation grounded in physics and data sciences. The model’s success illustrates the maturation of machine learning techniques in handling complex geophysical phenomena and elevates the forecasting community’s ability to exploit AI for operational meteorology. Dr. Brown emphasized that the ensemble model optimizes the balance between computational efficiency and predictive skill, a critical factor for future developments in climate and weather services.

The deployment of AIFS ENS is also an integral component of ECMWF’s broader engagement with open-source AI forecasting frameworks, particularly the Anemoi system developed collaboratively with Member States. The Anemoi framework provides an open platform for training and evaluating AI forecasting models, offering transparency and extensibility needed for widespread community contributions and evaluation. This ongoing co-development aims to foster cutting-edge AI methodologies while ensuring quality control and adaptability in various meteorological contexts.

Florian Pappenberger, ECMWF’s Director of Forecasts and Services, elaborated on the complementary relationship between the AI-based AIFS models and the traditional Integrated Forecasting System (IFS). By offering multi-faceted forecast products, ECMWF enables users to select the most appropriate outputs according to their operational demands. The continuation of 24/7 operational support further solidifies ECMWF’s commitment to integrating AI models like AIFS ENS into the mainstream meteorological workflow, while fostering continual improvements informed by real-world application feedback.

Moreover, the energy efficiency of AIFS ENS is a pivotal milestone amidst increasing awareness of sustainability within computational sciences. By drastically cutting the resource-intensive nature of ensemble forecasting, the AI-driven approach aligns with global goals to reduce carbon footprints in scientific computing. This breakthrough suggests the potential for scaling weather forecasting infrastructure without proportional increases in environmental impact, a crucial consideration for the global climate science community.

In summary, the unveiling of the AIFS ENS model by ECMWF signifies a paradigm shift in medium-range weather forecasting. Integrating AI into ensemble methodologies amplifies prediction accuracy and operational efficiency while fostering international collaboration through open-source development. Future advancements alongside hybrid systems promise to elevate both spatial resolution and forecast fidelity, reaffirming ECMWF’s role as a global pioneer at the confluence of meteorology and frontier data science.

Subject of Research: Not applicable

Article Title: (Information not provided)

News Publication Date: Tuesday 1st July 2025

Web References:

• ECMWF Overview of Ensemble Forecasting: https://www.ecmwf.int/en/about/media-centre/focus/2017/fact-sheet-ensemble-weather-forecasting
• Anemoi Framework Award: https://www.emetsoc.org/ems-technology-achievement-award-2025-for-anemoi/

References: (No specific references aside from web links)

Image Credits: ECMWF 2025

Keywords: Artificial intelligence, Atmospheric science