Algorithmische Optimale Steuerung - CO2-Aufnahme des Meeres

Use of Articifial Neural Networks for the acceleration of climate simulations

We want to extend existing work in using neural networks for the acceleration of computationally expensive climate simulations.

The following applications have already started previously in BSc/MSc and research work:

  • Approximation of steady annual cycles in marine ecosystem models (combination of ocean flow and marine biogeochemistry).
  • Approximation of one or several time steps of ocean circulation.
  • "Inverting" the simulation models, i.e., detecting climate model parameters from given simulation output.

Climate simulations are computationally very challenging. In the recent past, researchers started to use neural netowrks to accelaerate the simulation.

We therefore want to extend these work. This is a current research topic.

We have used different network structures, sparse training algorithms and evolutionary ideas for the choice of network hyperparameters.

We used Keras/TensorFlow and can provide access to high performance computers.

We look for students interested and (in the ideal case) experienced with neural networks and the above software frameworks.