The increasing availability of large datasets in the geosciences induces a trend to utilize machine learning algorithms and data-driven modelling as an alternative to process-based approaches. Though often successful in fitting observations, they are plagued by at least three issues: lack of interpretability; violation of conservation laws or more general lack of physical consistency; and the high dimensionality of data sets combined with lacking domain knowledge, implying the risk to include causally unrelated variables into the ML models.
Recent developments aim at bridging the gaps between “processes and data” and to mitigate these issues. Notably, Interpretable Machine Learning is a developing field, embracing efforts to render model predictions comprehensible. To tackle the second issue, a framework to implement physical constraints or other relevant domain knowledge into ML models is emerging, provisionally coined physics-aware ML modelling. Finally, the data-driven causal inference and causal network construction are promising avenues to constrain predictor selection and connect to process knowledge.
For this session, we invite contributions from the biogeosciences dealing with one or more of the following topics:
- Application of interpretable ML methods to ecosystem datasets
- Combinations of process-based models and ML methods
- Implementation of physical constraints into ML approaches, e.g. physics-aware regression
- Identification of drivers using causal inferences / causal networks

Knowledge domains relevant for the session include, but are not limited to, carbon and nitrogen cycles in terrestrial ecosystems, hydrology, soil processes, remote sensing, paleobiogeosciences, or extreme events.