Facilitating ab initio configurational sampling of multicomponent solids using an on-lattice neural network model and...

Area of Science:

• Materials Science
• Computational Chemistry
• Solid-State Physics

Background:

• Predicting configurational properties of multicomponent crystalline solids is computationally challenging.
• Traditional methods often require extensive structural relaxation, increasing computational cost.

Purpose of the Study:

• To propose an ab initio configurational sampling scheme for crystalline solids using neural network potentials (NNPs).
• To bypass the need for structural relaxation in lattice configuration problems.

Main Methods:

• NNPs trained to predict energies of relaxed structures from disordered lattices, not continuous coordinates.
• Active learning employed to generate a thermodynamically relevant training set.
• Demonstration on spinel oxides (MgAl2O4, ZnAl2O4, MgGa2O4) to calculate A/B site inversion.

Main Results:

• Successfully demonstrated the scheme on three spinel oxides.
• Calculated the temperature dependence of A/B site inversion.
• Showcased the potential to bypass structural relaxation in NNP applications.

Conclusions:

• The proposed scheme offers an alternative to conventional methods like cluster expansion for complex systems.
• Applicable to multicomponent bulk and interface systems relevant to technological applications.
• Enables efficient configurational sampling in crystalline solids.