In many fields of science, and particularly in environmental science, process models have been developed to represent our knowledge of the mechanisms underlying fluxes and states. We hypothesize that such process knowledge can substantially improve purely data-driven neural networks in small data settings. Thus, we want to expand approaches for process models in neural networks by augmenting neural networks with existing biophysical process knowledge and using explainable AI to improve process models. We will specifically consider augmentation data by simulations from a process model, and direct uses of process models by regularization, post-processing of process-model output with a neural network, or by embedding the process model into a series of neural networks to combine process model fitting with prediction optimization.
(supervised by PI Bödecker)
Network architecture search across model designs, design comparison, new designs, and explainable artificial intelligence techniques.
(supervised by PI Dormann)
Implementing and operationalizing the use of process models, new designs, and explainable artificial intelligence techniques.