There are in very general term three important ingredients when fitting exchange correlation functionals for density functional theory. One needs to have good and representative benchmark systems of adequately high quality, one needs a very flexible functional form, and finally one needs a good algorithm for selecting a good fit which is not an overfit to the benchmark systems.

We work on all three aspects of functional fitting. We design very flexible functional forms to be fitted with an emphasis on including van der Waals interactions. We work on establishing benchmarks to be used in catalysis and surface science, for example by carrying out higher-level calculations for adsorbate-surface systems. We devise machine learning algorithms in order to avoid overfitting of the functionals. A special feature of the functionals we construct in the so-called "BEEF"-family is that they allow the user to estimate uncertainties in the electronic structure simulations.