The aim of this WP is to learn model errors from training data. In inverse problems it is usually considered imperative to have an accurate forward model of the underlying physics. Nevertheless, such accurate models can be computationally highly expensive due to possible nonlinearities, large spatial and temporal dimensions as well as stochasticity. Thus, in many applications approximate models are used in order to speed up reconstruction and to comply with hardware and cost restrictions. Therefore the introduced approximation errors need to be taken into account when solving ill-posed inverse problems.