The
accurate description of large molecular systems has triggered
the development of new computational methods. Due to the computational
cost of modeling large systems, the methods usually require a trade-off
between accuracy and speed. Therefore, benchmarking to test the accuracy
and precision of the method is an important step in their development.
The typical gold standard for evaluating these methods is isolated
molecules, because of the low computational cost. However, the advent
of high-performance computing has made it possible to benchmark computational
methods using observables from more complex systems such as liquid
solutions. To this end, infrared spectroscopy provides a suitable
set of observables (i.e., vibrational transitions) for liquid systems.
Here, IR spectroscopy observables are used to benchmark the predictions
of the newly developed GFN2-xTB semiempirical method. Three different
IR probes (i.e., N-methylacetamide, benzonitrile,
and semiheavy water) in solution are selected for this purpose. The
work presented here shows that GFN2-xTB predicts central frequencies
with errors of less than 10% in all probes. In addition, the method
captures detailed properties of the molecular environment such as
weak interactions. Finally, the GFN2-xTB correctly assesses the vibrational
solvatochromism for N-methylacetamide and semiheavy
water but does not have the accuracy needed to properly describe benzonitrile.
Overall, the results indicate not only that GFN2-xTB can be used to
predict the central frequencies and their dependence on the molecular
environment with reasonable accuracy but also that IR spectroscopy
data of liquid solutions provide a suitable set of observables for
the benchmarking of computational methods.