GeoMod4Future: Sustainable modeling for the geosciences

Modeling in the geosciences requires a change of perspective. It is a valuable tool for answering urgent questions, e.g. with respect to climate change or flood hazards. However, models are typically built for a single purpose, and this limits their range of applicability. By building a model along a predefined impact chain to the desired forecast output, such models are not equipped to address new questions arising from changing environmental or political conditions. We could significantly save computing resources and expand knowledge more effectively if we made existing models fit for new questions. To achieve this, a model must be structured and trained in a fundamentally different way. While traditional models only learn the conditional probability distribution of the target variable given the input variables, we propose network-like models that - inspired by recent methods from machine learning - link all system-relevant data types: neural stars. 

Through a novel multidirectional training, the neural stars learn the joint distribution of all data and establish a physically consistent and interpretable core. Such a model uses the existing data in an optimal way, learns more about the natural system, and should thus also be able to answer new, system-related questions without the need to be extended or retrained. For this innovative goal, we need three new methodological developments through fundamental research: 1) Probabilistic neural stars by superimposing conditional variational auto-encoders; 2) Multidirectional training instead of a rigid causal direction; 3) Smart selection layer inspired by reservoir computing to establish relationships from the learned abstract core to derived physical states/processes in the real world. In short: We deviate from the common modeling philosophy of "fit for purpose" for good reasons. Our vision is an open-purpose model, fit for many situations, that can honestly reveal its knowledge (or uncertainty) about previously unseen questions. We will investigate how efficiently this can be implemented. The aim of the 1.5-year project duration is a proof-of-concept by applying the developed methods to a real hydrological data set.