Artificial Intelligence (AI) is rapidly transforming the climate and weather forecasting landscape, as well as the assessment and management of natural hazard risks. With advancements in machine learning (ML) techniques, particularly sophisticated deep learning (DL) models, and the increasing availability of large datasets, AI-based technologies are receiving growing attention for their potential to revolutionize early warning systems. The AML group, leveraging cutting-edge machine learning methods, is focused on enhancing the accuracy and reliability of weather forecasts across various time scales, from nowcasting to sub-seasonal predictions. This work holds immense value for industries dependent on precise weather information, such as agriculture, transportation, and disaster management. By pushing the boundaries of traditional forecasting techniques, their efforts enable more accurate predictions of weather events, improving preparedness and response across a broader range of time horizons.
Rainfall to Streamflow Modelling
Adapting to climate change implies the intelligent management of water resources and waterway infrastructure. We operate a model that predict the amount of water in rivers, by integrating the weather forecast with location information from a variety of maps. This information can be leveraged directly generate local flood warnings or plan water way logistics. The model is flexible and can also be adapted to other application cases such as prediction soil moisture for agriculture or creating forest fire hazard warnings.
Downscaling
Downscaling in climate applications refers to the process of translating coarse-resolution outputs from global climate models (GCMs) or weather forecasts into high-resolution data suitable for local or regional analysis. Since GCMs typically operate on grid scales of tens to hundreds of kilometers, they lack the spatial detail necessary for impact assessments in sectors like agriculture, water management, and disaster response. There are two main approaches: dynamical downscaling, which uses regional climate models nested within GCMs, and statistical (or empirical) downscaling, which learns relationships between large-scale and local variables from historical data. Recently, machine learning and generative models like diffusion models have emerged as powerful tools for statistical downscaling, offering flexible, probabilistic frameworks to generate fine-scale climate data. These methods are especially valuable for applications requiring realistic, high-resolution scenarios from low-resolution forecasts, such as subseasonal-to-seasonal (S2S) prediction.
The Applied Machine Learning (AML) group has done extensive work on downscaling with diffusion-based models. In the Context of S2S predictions, DiffScale [1], a multimodal diffusion model was developed. DiffScale is a conditional diffusion model designed for dynamic downscaling of subseasonal-to-seasonal (S2S) weather forecasts. It learns to transform coarse, low-resolution climate predictions into high-resolution outputs by modeling the transition as a score-based diffusion process. The model leverages spatial and temporal patterns in the input data to generate fine-scale details consistent with physical dynamics. By conditioning on low-resolution forecasts, DiffScale produces realistic, high-fidelity weather scenarios at finer scales. This approach offers a probabilistic, data-driven alternative to traditional downscaling methods, improving resolution without explicit physical modeling.
Deep learning-based post-processing for S2S forecasts of soil moisture
Subseasonal to Seasonal (S2S) forecasting bridges the gap between medium-range weather forecast (up to ~10 days) and seasonal predictions (3–6 months), playing a critical role in planning for sectors like energy, water management, and agriculture. However, these forecasts are challenging due to the limited predictability between short-range weather forecasts (around 10 days) and seasonal climate projections. The accuracy of S2S forecasts is constrained by both the inherent unpredictability of the Earth’s chaotic systems and errors in numerical models, which are influenced by various uncertainties, including those related to initial conditions and model parameterizations. These uncertainties result in both systematic and random errors that quickly amplify over time.
Accurate subseasonal soil moisture forecasts are essential for improving predictions of flash droughts, which, unlike traditional droughts, develop and intensify over shorter timescales. We use a hybrid modeling approach that combines numerical weather predictions with deep learning (DL) models. This approach not only corrects errors in the subseasonal forecasts but also enhances their spatial resolution, providing more detailed and reliable predictions.
Nowcasting
Nowcasting is the very short‑term forecasting of atmospheric conditions over the next minutes to a few hours. It is critical for issuing timely severe‑weather warnings including tornadoes, flash floods and hailstorms. Precipitation and extreme‑weather nowcasting are particularly challenging within the weather domain due to the highly non‑linear and chaotic nature of underlying processes, yielding low predictability.
As many phenomena are not well captured by conventional methods, deep learning poses a promising direction to deal with the vast amounts of data available in the weather domain. We develop multi-modal generative models to enable accurate uncertainty quantification, integrating data sources such as radar reflectivity, high‑resolution satellite imagery and physics‑based variables such as wind velocity, humidity and temperature.
Publications