HD res photo
Background
Understanding and accurately simulating the terrestrial water cycle is essential for predicting hydrological extremes, managing water resources, and assessing the impacts of climate change. Land-surface and hydrological processes also help constrain the atmosphere and, through their longer memory, can enhance the reliability of forecasts at longer lead times. Physically based hydrological and land-surface models underpin many operational forecasting systems, yet they often struggle to represent nonlinear interactions, parameter uncertainty, and heterogeneous and scale-dependent land-surface processes—particularly at the spatial and temporal scales required for real-time decision-making.
Recent advances in machine learning (ML) provide complementary pathways to improve hydrological and land-surface prediction. These include fully data-driven ML models trained directly from observations, hybrid approaches that combine ML with process-based models, and physics-informed methods that embed physical constraints into learning algorithms. Together, these approaches offer new opportunities for scalable, computationally efficient, and observation-driven prediction systems suitable for operational and pre-operational use. In addition to improving predictive components, ML also enables advances in data assimilation, which can help integrate emerging satellite and in situ datasets into forecasting systems more effectively and at reduced computational cost.
Workshop overview
The workshop brought together experts in machine learning, hydrology, and land-surface modelling to discuss the development and use of ML-based and ML-enhanced forecasting systems. It focussed on datasets, methods, applications, and evaluation strategies for machine-learning approaches, covering fully data-driven, hybrid, and physics-informed models. Contributions spanned both global and regional perspectives, addressing ML-based prediction of streamflow, floods, soil moisture, and key land-surface processes, with an emphasis on robustness, scalability, and operational relevance.
Discussions addressed scientific and operational challenges such as data availability and latency, robustness under extremes and non-stationarity, uncertainty quantification, evaluation and benchmarking, and the transferability of ML models across regions, climates, and time scales. By explicitly linking observation-driven ML approaches with operational requirements, the workshop aimed to inform the development of the next generation of hydrological and land-surface forecasting systems.
Expected outcomes
-
A synthesis of emerging ML methods and use cases for hydrological and land-surface modelling.
-
Community recommendations for data standards, benchmarking, and evaluation practices.
-
Identification of key research priorities and future directions.
-
Strengthened collaboration between the hydrology, land-surface, and ML communities, including ECMWF Member States.
Format
This workshop was designed exclusively for in-person attendance. There were no virtual or online attendance options available. We welcomed attendees on site and engaged face-to-face throughout the event.
The workshop ran over 2.5 days and was structured around the full pipeline of ML for land surface modelling. Day 1 focussed on novel ML methods and applications for land surface processes, with sessions on new methods and ML applications. Day 2 tackled "Closing the Reality Gap" — covering datasets and observations, and evaluation and benchmarking, with discussions on data assimilation challenges and whether we need ML-specific evaluation approaches. Day 3 was a half-day focused on moving to operations, with a discussion on what's still needed from both modellers and users.
Each session combined keynotes, submitted short talks, themed discussions and debates (e.g. "What is the role of physical models in land surface ML?"), and working group breakouts to scope concrete outputs. There was also a poster session, an ECR panel on the future of scientists in the field, and a final synthesis session to capture research priorities and next steps.
Attendance
The workshop took place at ECMWF's headquarters in Reading, UK over 2.5 days from 3 to 5 of November. Participation was open to researchers, operational scientists, and practitioners working in hydrology, land-surface modelling, Earth system science, and machine learning. Priority was given to registrants from organisations in ECMWF's Member States.
Whilst there were no registration fees, we were unable to provide funding to support attendance. Participants were expected to cover their own travel, meal, and accommodation costs.
Timeline
5 February 2026: Registration and abstract submission open
26 June 2026: Abstract submission deadline
31 July 2026: Notification of abstract acceptance
4 September 2026: Registration closes