AI4S2S

Weather extremes like heatwaves and droughts severely impact society and effective early warnings require forecasts at seasonal to sub-seasonal (S2S) timescales. S2S-predictability is region-dependent and intermittent in time, creating spatio-temporal ‘windows of predictability’ within a largely unpredictable future. Massive climate datasets and machine learning provide huge opportunities to strategically target these windows. However, predictability by itself is
insufficient: Trustworthy data-driven forecasts require (1) transparent and reproducible analyses, (2) best practices in model verification, and (3) understanding of the sources of predictability. In project AI4S2S, we built open-source, high-level python libraries that provide an interface between artificial intelligence and expert knowledge, to boost predictability and physical understanding of S2S processes. These python libraries provide high-level functionality for scientists who want to use AI for S2S forecasting: The ‘Lilio’ library presents a calendar generator for machine learning with timeseries data; the ‘s2spy’ library enables users to integrate expert knowledge and artificial intelligence to S2S forecasting. These libraries form the core libraries of the AI4S2S data-pipeline that enables to build specific recipes for S2S forecasting. In the final year, we explored ways of including explainable AI (XAI) functionality
in this pipeline, including the eScience Center’s DIANNA framework. Over the last few years there has been a true revolution in the use of AI to climate and weather forecasting problems, with new AI methods including Atmosphere Foundation Models becoming available. This has generated new opportunities during the lifetime of project AI4S2S. In particular, we invested in bridging and uniting different sub-disciplines by organizing 2 workshops: 1) The five-day
Lorentz workshop ‘Boosting (sub) seasonal forecasts with Explainable AI’ (Leiden, Sept 2022) united climate scientists, data scientists, seasonal-forecast experts and XAI experts to improve sub-seasonal to seasonal (S2S) predictions of droughts in the Horn of Africa together with local stakeholders. 2) The five-day ‘AI in Weather & Climate’ (Texel, July 2025) brought together climate scientists, seasonal forecasting experts, AI experts, data scientists and software developers to build a stronger Dutch AI-Climate community focused on building better and
more sustainable open source code. In this last workshop the newly formed community network AIMET-NL (AI for Meteorology and Climate – a Dutch community) was founded. AIMET-NL aims at uniting and strengthening the Dutch community working on AI challenges in climate and currently unites 10 institutes across the Netherlands.

The AI4S2S project was supported by the OEC2021 and additionally by an internal software sustainability project titled "Artificial Intelligence for S2S scientists".