Figure 1.

a), b) A one‐dimensional energy landscape (blue) is sampled at some select points (grey points) and a GPR model (yellow) is established. a) With an evolving population, locally optimal data points (arrows) being sufficiently different will constitute the population and not necessarily represent all data. b) With a clustering‐based sample scheme enforcing locally optimal data points (arrows) to be drawn from different clusters (colored points) a more representative sample of data is obtained. Using such a sample thus has potential to evolve more exploratively compared to a population based search. In this example, descendants from the data drawn from the red cluster are expected to more easily evolve into the right part of the energy landscape. c) Illustration of the sample generation scheme based on data from a GOFEE search for 2D Sn₃O₆ nano‐clusters with N _sample=5. First, the current set of DFT evaluated structures are represented in a feature space and are subsequently clustered using the k‐means++ algorithm, to identify groups of related structures. The sample is then created by selecting the most stable structure from each group.