De Novo Inverse Design Superhard C–N Compounds via Global Machine Learning Interatomic Potentials and Multiobjective Optimization Algorithm

A major challenge in the field of superhard materials
is the identification
of compounds with a hardness exceeding that of diamond. In this study,
we developed a variable-composition inverse material design (VC-IMD)
approach for designing C–N superhard materials. In this approach,
an improved multiobjective optimization algorithm is introduced, utilizing
structure similarity constraint to prevent convergence toward local
minima. Combined with active learning, it trains global machine learning
interatomic potentials (g-MLIPs) while exploring
target materials. By comparing several g-MLIPs and
selecting the best, the resulting g-MLIPs achieved
reasonable precision within three iterations. Through multiple searches,
38 novel and stable C–N superhard materials not present in
major computational materials databases were identified. Notably,
the material C3(P6422) with
a hardness of 97.4 GPa was discovered, potentially exceeding that
of diamond (94.0 GPa). This approach provided a new pathway for materials
design with target properties.