Evolutionary algorithm
This paper uses machine learned atomic potential to guide the rattle and crossover operations in an EA run, thus accelerating the structure search in SnO2(110)-(4x1) reconstruction, anatase TiO2(001)-(1x4) ridge reconstruction and quinoline(C9H7N). The atomic potential is defined as the energy difference between one structure (n atoms) and the same structure lacking one atom (n-1 atoms). $$E _ {atomic} = E(\bold{R} _ n) - E(\bold{R} _ {n-1})$$ A low $E_{atomic}$ means this atom is in a stable environment. Thus the atom with high $E_{atomic}$ should have high probabilities to be displaced in the rattle operation and removed in crossover operation.