Deep learning and Process Understanding for Data-Driven Earth System Science

For a better understanding of the Earth system we need a stronger integration of observations and (mechanistic) models. Classical model-data integration approaches start with a model structure and try to estimate states or parameters via data assimilation and inverse modelling, respectively. Sometimes, several model structures are employed and evaluated, e.g. in Bayesian model averaging, but still parametric model structures are assumed. Recently, Reichstein et al. (2019) proposed a fusion of machine learning and mechanistic modelling approaches into so-called hybrid modelling. Ideally, this combines scientific consistency with the versatility of data driven approaches and is expected to allow for better predictions and better understanding of the system, e.g. by inferring unobserved variables. In this talk we will introduce this concept and illustrate its promise with examples on biosphere-atmosphere exchange, and carbon and water cycles from the ecosystem to the global scale.

Reichstein, M., G. Camps-Valls, B. Stevens, M. Jung, J. Denzler, N. Carvalhais, and Prabhat. "Deep Learning and Process Understanding for Data-Driven Earth System Science." Nature 566, no. 7743 (2019): 195-204. https://doi.org/10.1038/s41586-019-0912-1.