WMO AI Webinars

The AI Weather Quest: Framework and Participation over the First Year

Abstract: A look back at the Quest’s first year, from its forecasting framework to the global community of participating teams.

presentation(pdf),

recording

Scientific Insights and Rankings

Abstract: Key scientific insights from the first year of the Quest, together with the JJA 2026 rankings and overall performance highlights.

presentation(pdf),

recording

Fengshun: Advancing AI-Driven Subseasonal-to-Seasonal Prediction

Abstract: The CMAandFDU team has developed the Fengshun family of subseasonal-to-seasonal (S2S) prediction models which participated in the AI Weather Quest: Fengshun, FengshunAdjust, and FengshunHybrid. Fengshun is a lightweight, purely data-driven model that directly predicts weekly probabilistic quintiles of near-surface temperature, precipitation, and mean sea-level pressure. FengshunAdjust is a purely data-driven ensemble model that combines three complementary AI modeling paradigms—Transformer-based, Diffusion-based, and GroupVAE/Mamba-based models—to leverage their respective strengths and improve forecast skill and robustness across variables, regions, and lead times. FengshunHybrid further extends this ensemble strategy by combining three data-driven models with operational dynamical S2S prediction models, thereby exploiting the complementary strengths of AI-based and physics-based forecasting.

The Fengshun family has demonstrated strong performance in the AI Weather Quest. FengshunHybrid won the third season of the competition, while in the ongoing fourth season, FengshunAdjust is currently leading the competition, demonstrating the strong potential of purely data-driven ensemble approaches for S2S forecasting. This presentation will introduce the design of the Fengshun family, discuss the complementary strengths of these models, and offer perspectives on the future of operational S2S prediction in the AI era.

Bo Lu (Co-Team Leader, China Meteorological Administration)

Zesheng Dou (Developer of the Fengshun Foundation Model, China Meteorological Administration)

Yang Zhao (Developer of the Fengshun Ensemble Algorithms, China Meteorological Administration)

presentation(pdf),

recording

Advancing Subseasonal Forecasting with Machine Learning

Abstract: In this presentation, the MicroEnsemble team participating in the AI Weather Quest introduces probabilistic bias correction (PBC), a machine learning framework that substantially reduces systematic error by learning to correct historical probabilistic forecasts. Applied to ECMWF's leading dynamical and AI models, PBC doubles the AI Forecasting System's modest subseasonal skill and improves the operationally-debiased dynamical model on 91% of pressure, 92% of temperature, and 98% of precipitation targets. These probabilistic skill gains translate into more accurate prediction of extreme events and have the potential to improve agricultural planning, energy management, and disaster preparedness in vulnerable communities.

presentation(pdf),

recording

How Far Does Forecast Bias-adjustment Take You?

Abstract: The CLINT team (Climate Data Factory, CMCC Foundation) reached the top five of the AI Weather Quest without the AI it set out to build. The skill comes from using and bias-adjusting ECMWF’s ensemble and combining lagged forecasts, an approach that also beats the operational dynamical benchmark. Our deterministic data-driven model, a multi-encoder residual U-Net pre-trained on a 2000-year paleoclimate simulation and fine-tuned on ERA5, entered only in the final two weeks of the competition, where it is used to shift the distribution of the probabilistic dynamical forecast. We give an overview of both approaches and their achievements: the added value obtained from the bias-adjustment and lagged-ensemble route, and the first results from the hybrid approach.

presentation(pdf),

recording

What Comes Next for the AI Weather Quest

Abstract:  An overview of the Quest’s next phase, including new forecast variables and upcoming opportunities to participate.

presentation(pdf),

recording

MLP Correction of Solar Radiation Forecasts from NWP Models in Argentina

Abstract: This presentation introduces a machine learning approach based on Multi-Layer Perceptrons (MLP) to improve solar radiation forecasts generated by Numerical Weather Prediction (NWP) models. Developed through collaboration between the Argentine National Meteorological Service (SMN) and the National University of Salta (UNSa), the AI-based correction enhances forecast accuracy, providing more reliable climate services for solar energy generation, power system operations, and renewable energy planning.

AI-Based Downscaling of Climate Projections for Renewable Energy Atlases in South America

Abstract: This presentation introduces a participatory framework for developing national renewable energy atlases that combine observations, climate projections and artificial intelligence to support local decision-making. Building on the WMO Renewable Energy Atlas work in Argentina, Chile, Peru and Ecuador, the approach applies AI-based downscaling to transform global climate projection data into more locally relevant information for wind, solar and hydropower resource assessment. The presentation highlights how the co-development of climate services with NMHSs and energy-sector stakeholders can strengthen long-term renewable energy planning and climate-resilient investment decisions across South America.

Operational subseasonal to seasonal (S2S) prediction of drought using ML Downscaling

Abstract: Drought has significant impacts on New Zealand’s export-driven economy. In 2022, the Ministry for Primary Industries, in partnership with NIWA, launched a 35-day drought forecast based on the New Zealand Drought Index (NZDI), to mitigate these impacts.
This system uses coarse-resolution forecast inputs and applies AI-based downscaling to capture New Zealand’s complex terrain, effectively refining forecasts from 50 km to 5 km resolution. The result is a tool that enables farmers and other primary producers to make more informed, timely decisions, helping to reduce financial losses and manage stock more effectively.
This talk will outline the development and operationalisation of this forecasting system, as well as its ongoing application, assessment and impact.

Integrating AI in Seamless Prediction from Nowcasting to Medium-Range Weather Forecast

Abstract: The Hong Kong Observatory (HKO), serving as the WMO Regional Specialized Meteorological Centre (RSMC) for Nowcasting, has been actively developing and operationalizing artificial intelligence (AI) applications across the full spectrum of weather prediction timescales. This presentation outlines HKO's progressive integration of AI and machine learning (ML) techniques — from rainstorm nowcasting to medium-range forecasting — for advancing seamless weather prediction framework.

In the nowcasting domain, HKO has continuously enhanced deep learning precipitation forecast models, and has integrated a suite of different deep learning frameworks into the operational SWIRLS nowcasting system.  To extend nowcast lead time with broader geographical coverage for supporting WMO Members in southeast Asia, the AI nowcast based on simulated reflectivity of multispectral imagery of geostationary satellite has been implemented in the RSMC for Nowcasting website in December 2025.

AI techniques have been widely applied in enhancing short-range to medium-range forecasts.  Besides running several AI weather prediction (AIWP) models in-house and using available AIWP data products such as ECMWF AIFS, post-processing techniques have been used to enhance multi-model ensemble products with a view to enhancing location-specific forecasts in Hong Kong.  Evaluations demonstrate that the integration of post-processed forecasts from AIWPs with NWPs can further improve the automatic location-specific weather prediction out to 2 weeks ahead.

Together, these developments illustrate a paradigm shift towards an integrated AI-augmented seamless prediction chain. Ongoing developments - such as concept of “AI on AI” to further leverage benefits from AI nowcast and medium-range forecast models for improving high-impact weather predictions, as well as challenges ahead will be introduced.

Advancing Nowcasting with Deep Learning techniques (ANDeL) for West Africa

Abstract: Accurate short-term rainfall prediction (0–6 hours lead time) remains a critical challenge across much of Africa, where sparse observational networks and the limitations of conventional numerical weather prediction systems hinder the representation of localized convective processes. The Advancing Nowcasting with Deep Learning techniques (ANDeL) project leverages deep learning architectures (convolutional LSTM and attention-based models) to predict the spatio-temporal evolution of rainfall using multi-source datasets [satellite-derived precipitation (IMERG) and reanalysis (ERA5)]. Initial model training was conducted using IMERG to establish a robust baseline; however, due to its latency (~4 hours), current operational testing employs Rain-over-Africa (RoA) data, which provides low-latency, high spatio-temporal resolution inputs suitable for near-real-time applications. The framework incorporates transfer learning and adaptive fine-tuning to enable efficient deployment across diverse regions, while maintaining a strong operational focus on low-compute environments. 

WAS-NextGen: An Objective Multi-Method Framework for Seasonal Climate Forecasting in West Africa

Abstract: Seasonal climate forecasting in West Africa has traditionally relied on consensus-based methods that lack reproducibility and high-resolution detail. This presentation introduces WAS-NextGen, an automated, multi-method framework that integrates machine-learning-calibrated multi-model ensembles (SV–ML–CMME), statistical–dynamical CCA calibration, lagged-predictor components, and analogue-year methods. Implemented via the open-source Python package wass2s, the system ensures a fully reproducible workflow from data acquisition to probabilistic mapping.

Skilful long-lead nowcasting with NowAlpha in operations

Integration of AI/ML into operational weather and environmental forecasting systems at ECCC

Environment and Climate Change Canada (ECCC) is integrating AI/ML into operational weather and environmental forecasting systems. This presentation will summarize the progress on the AI-physics hybrid Global Deterministic Prediction System with Spectral Nudging (GDPS-SN) and describe its path to operationalisation. The presentation will also describe progress made on the AI-based PARADIS weather model and feature other AI/ML projects for weather and environmental forecasting currently underway at ECCC.