Dear all,

I have some general questions about the usage of the cluster expansion. I will use the enumeration algorithm (enumlib) that is also implemented in icet to create a set of structures based on an input structure (in my concrete case, I want to find the antiferromagnetic ground state configuration). Given this set of structures, I have to choose a subset of structures to create the cluster expansion model, right?

First, the structures in the subset to create the cluster expansion model need to be relaxed, right? But the model is afterwards also valid for the unrelaxed structures, otherwise CE would not give a computational advantage.

Second, what are the limits of one CE model? When I create a CE model based on a set of structures that was obtained by enumerating a specific input structure, is it then possible to also use the trained model to predict the properties of another structure that was not related to the “original” structure? So, the question is, when I have two structures that consist of Ni and O atoms and are different concerning their crystal structure, would it be possible to train a CE model just based on the enumerated set of the first structure and to use this model then to predict the properties of the second structure (that is not related to the enumerated set of the first one)?

Thanks a lot in advance.