An invertible, invariant crystal representation for inverse design of solid-state materials using generative deep...

Area of Science:

• Materials Science
• Computational Chemistry
• Solid-State Physics

Background:

• Deep learning has advanced molecular design using representations like simplified molecular-input line-entry system (SMILES).
• Designing solid-state materials with specific properties remains challenging due to the lack of suitable crystal representations.
• Existing methods lack invertibility and invariance for crystal structures, hindering inverse design.

Purpose of the Study:

• To develop an invertible and invariant crystal representation for efficient materials design.
• To enable the inverse design of solid-state materials with targeted properties.
• To address the limitations of current crystal structure representations in computational materials discovery.

Main Methods:

• Developed a simplified line-input crystal-encoding system (SLICES), a string-based representation.
• Ensured SLICES satisfies translational, rotational, and permutational invariances by encoding compositional and topological data.
• Validated SLICES through a reconstruction routine on diverse crystal structures.

Main Results:

• SLICES achieved 94.95% reconstruction accuracy for over 40,000 diverse crystal structures, demonstrating high invertibility.
• The representation guarantees invariances by focusing on compositional and topological information.
• Successfully applied SLICES to the inverse design of direct narrow-gap semiconductors for optoelectronics.

Conclusions:

• SLICES provides an effective, string-based, invertible, and invariant crystal representation.
• This novel representation facilitates in silico materials discovery and inverse design.
• SLICES shows significant promise for advancing the design of functional solid-state materials.