A FAIR Protocol for Hybrid Models and Data in Hydrology

Hybrid models, which combine physics and machine learning (ML) based models, are becoming increasingly popular in hydrology and the broader Earth Science community due to their potential for improved prediction and process representation. However, hybrid models pose unique challenges to open research practices, including the widely accepted FAIR principles. Unlike physics-based models, the reusability of hybrid models is hindered by the integration of ML models which dynamically change with training data. Furthermore, existing model and data repositories are not designed to host hybrid models which contain code, ML models, and associated training data. To address these challenges, FRAME will collaboratively design, implement, and test a standardised FAIR protocol tailored for hydrological hybrid models. The protocol will consist of coding standards for interoperability between different model components, a unified metadata specification accounting for different types of physics and ML-based models, and a python package leveraging existing model and data repositories widely used in the hydrology (HydroShare) and ML (DLHub) communities to share and retrieve hybrid models. To ensure wider and long-term impact of the project beyond its lifetime, the developed protocol will be actively used and improved by participating groups in the ETH Domain and Europe and will ultimately be transitioned to a community-driven protocol, inviting participation from the wider scientific community.