In this project we will be proposing a novel model evaluation framework for hybrid models which we define as any model found on the axis between fully physics-based and purely data-driven. Said framework will be rooted in information-theoretical concepts and methods. Initially the framework will be developed for the evaluation of hydrological models which simulate the transformation from rainfall to runoff in the water cycle. Traditionally, this has been modelled through a conceptual framework of the physical processes involved, such as evaporation and infiltration. More recently, data-driven approaches and particularly deep-learning models based on recurrent neural networks have gained popularity due to their ability to match and even outperform traditional conceptual models. Yet, these approaches have been criticized because currently they lack interpretability and a theoretical foundation for prediction under changing conditions. Physics-informed data-driven models, or hybrid models as we have previously defined them, have emerged in order to combine both modelling techniques and achieve model interpretability and prognostic skill. However, to date, there is no formal framework to evaluate and compare the quality of hybrid models and it is still unclear how to quantify and treat the prediction uncertainty of these hybrid approaches. The goal of the proposed project is therefore to set a new standard for diagnostic evaluation of hybrid models. Some of the expected benefits of this framework will be:
- Testable predictions that help advance science through improved system understanding.
- Guidance towards model improvement across the continuum of physics-based to data-driven models.
- Increased public acceptance of (hybrid) predictions through mapping of model entities and dynamics to real-world physical compartments, states and processes.
In the end, we will develop, test and demonstrate the proposed toolbox of information-theoretic methods on a specific case study of rainfall-runoff modelling. Nonetheless, our toolbox and evaluation framework will be applicable to hybrid models in a broad range of any scientific disciplines with societal importance.
|Project Number||PN 6A-4|
|Project Name||Unified diagnostic evaluation of physics-based, data-driven and hybrid hydrological models based on information theory|
|Project Leader||Anneli Guthke|
|Project Members||Manuel Álvarez Chaves, PhD Researcher|