Characterization of potential energy surfaces using machine-learning techniques

Machine-learned surrogate models are used to find minima, saddle points and reaction paths on potential energy surfaces of materials. These are used to characterize the energetics of chemical processes or the structure of materials on an atomistic scale. We build machine-learned models on-the-fly as more and more data from electronic structure calculations become available. Optimization algorithms are developed and adjusted to different ML approaches to deal with the inherent numerical noise in the data. The re-use of data as well as the combination of data sources with different cost and accuracy increases the computational efficiency. The goal is to gain as much knowledge about stationary points and optimal paths on the potential energy surface with as few electronic structure calculations as possible, since the latter are the computational bottleneck.