Space group informed transformer for crystalline materials generation.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Generative modeling of crystalline materials is computationally intensive.
• Existing methods struggle with the complexity of crystal space and symmetry.
• Data and compute efficiency are critical for large-scale materials discovery.

Purpose of the Study:

• To introduce CrystalFormer, a novel transformer-based autoregressive model for space group-controlled crystalline material generation.
• To leverage explicit space group symmetry for enhanced data and compute efficiency.
• To demonstrate CrystalFormer's applicability in structure initialization, element substitution, and property-guided materials design.

Main Methods:

• Developed CrystalFormer, a transformer architecture incorporating space group symmetry.
• Utilized the discrete and sequential nature of Wyckoff positions for prediction.
• Trained the model on a materials dataset to learn solid-state chemistry heuristics.

Main Results:

• CrystalFormer significantly reduces the complexity of crystal space for generative modeling.
• Demonstrated superior performance in symmetric structure initialization and element substitution compared to conventional methods.
• Showcased successful plug-and-play application in property-guided materials design, highlighting flexibility.

Conclusions:

• CrystalFormer offers a simple, general, and adaptable architecture for crystalline materials generation.
• The model effectively ingests chemical knowledge, enabling systematic exploration of materials space.
• CrystalFormer is positioned as a foundational model for a new era in materials discovery and design.