A generative model for inorganic materials design

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Functional materials are crucial for technological advancements in energy storage, catalysis, and carbon capture.
• Current generative models for materials design struggle with proposing stable crystals and satisfying multiple property constraints.
• Existing methods exhibit low success rates in generating novel and stable material structures.

Purpose of the Study:

• To introduce MatterGen, a generative model capable of designing stable and diverse inorganic materials.
• To enhance the generation of materials with specific chemical, physical, and electronic properties through fine-tuning.
• To overcome the limitations of existing generative models in terms of stability and property constraint satisfaction.

Main Methods:

• Development of MatterGen, a generative AI model for de novo materials design.
• Fine-tuning MatterGen to steer material generation towards desired property constraints (chemistry, symmetry, mechanical, electronic, magnetic).
• Evaluation of generated structures for novelty, stability, and proximity to energy minima compared to previous models.

Main Results:

• MatterGen generates novel and stable inorganic materials with over twice the success rate of prior models.
• Generated structures are significantly closer to the local energy minimum, indicating higher stability.
• Fine-tuned MatterGen successfully produced materials meeting specific property targets, with one synthesized material validating predicted properties within 20% of the target.
• Demonstrated ability to generate materials with desired chemistry, symmetry, and mechanical, electronic, and magnetic properties.

Conclusions:

• MatterGen represents a significant advancement in generative materials design, producing higher quality and more stable materials.
• The model's ability to be fine-tuned for diverse property constraints broadens its applicability.
• MatterGen shows promise as a foundational tool for accelerating the discovery of functional materials.