CrystalReasoner/Qwen2.5-3B-CrysReas-SpaceGroup
CrystalReasoner/Qwen2.5-3B-CrysReas-SpaceGroup is a Qwen2.5-3B-based large language model developed by CrystalReasoner, specifically fine-tuned for generating crystal structures. This model uses supervised fine-tuning (SFT) and reinforcement learning (RL) to generate crystal structures from natural language instructions, incorporating crystallographic and physical priors. It excels at property-conditioned crystal structure generation, providing lattice vectors, element types, and atomic coordinates.
Loading preview...
CrystalReasoner/Qwen2.5-3B-CrysReas-SpaceGroup Overview
This model, developed by CrystalReasoner, is an end-to-end large language model (LLM) framework built upon the Qwen2.5-3B architecture. Its primary function is to generate crystal structures based on natural language instructions, a specialized application within materials science and chemistry.
Key Capabilities
- Crystal Structure Generation: Translates natural language descriptions of bulk materials into detailed crystal structures.
- Property Conditioning: Incorporates specified material properties (e.g., bulk modulus) into the generation process.
- Reasoning Integration: Utilizes "thinking traces" to apply crystallographic and physical priors before generating atomic coordinates.
- Reinforcement Learning (RL): Employs RL with verifiable rewards to enhance the validity, stability, and property conditioning of generated structures.
- Output Format: Generates descriptions of lattice vectors, angles, element types, and atomic coordinates, which can be converted into
pymatgen Structureobjects.
When to Use This Model
This model is ideal for researchers and developers who need to:
- Generate novel crystal structures based on specific chemical formulas and desired physical properties.
- Automate the creation of crystal structure data from textual descriptions.
- Explore the design space of materials with property constraints.
It offers a unique approach to materials discovery by integrating advanced LLM capabilities with domain-specific reasoning and reinforcement learning for crystal structure prediction.