Generative Adversarial Networks for Crystal Structure Prediction.

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Accelerating the discovery of novel functional materials is crucial.
• Crystal structure prediction is a fundamental task in materials discovery.
• Generative models offer new avenues for exploring chemical space.

Purpose of the Study:

• To develop an efficient generative model for crystal structure prediction.
• To apply the model for discovering new Mg-Mn-O ternary materials.
• To evaluate the photoanode properties of predicted structures for high-throughput virtual screening (HTVS).

Main Methods:

• Utilized an inversion-free crystal representation based on unit cell and fractional atomic coordinates.
• Developed a generative adversarial network (GAN) for crystal structure generation.
• Performed theoretical evaluation of photoanode properties and calculated stability and band gap.

Main Results:

• Generated 23 new Mg-Mn-O ternary crystal structures.
• Predicted structures exhibited reasonable calculated stability.
• Identified materials with potential for photoanode applications.

Conclusions:

• Generative models provide an effective approach for exploring uncharted chemical spaces.
• The proposed generative HTVS framework accelerates the discovery of novel functional materials.
• This method expands beyond conventional substitution-based materials discovery.