Atom Classification Model for Total Energy Evaluation of Two-Dimensional Multicomponent Materials

: The tunable properties of materials originate from the variety of structures, however, it is still a challenge to give an accurate and fast evaluation of stabilities for screening numerous candidates. Herein, we propose an atom classification model(ACM) to describe the multicomponent materials based on the structural recognition, in which the atoms are classified to estimate the total energies. Taking twodimensional planar C1-xBx and C1-2x(BN)x as examples, we have found that the test error of total energies is about 3 meV per atom. Notably, the distributions of classified atoms demonstrate the evolution of configurations as a function of temperature, providing a more clear picture of phase transition. In addition, our method is universal which can be flexibly extended to the bulk structures with more components.