CrystalReasoner/Qwen2.5-3B-CrysReas-ThermalExpansion

CrystalReasoner/Qwen2.5-3B-CrysReas-ThermalExpansion Overview

This model, developed by CrystalReasoner, is an end-to-end large language model (LLM) framework designed for generating crystal structures based on natural language instructions. It leverages a Qwen2.5-3B base and is specialized through a multi-stage training process.

Key Capabilities

• Crystal Structure Generation: Generates detailed descriptions of lattice vectors, angles, element types, and atomic coordinates for crystal structures from textual prompts.
• Property Conditioning: Incorporates physical and crystallographic priors into the generation process, allowing for property-conditioned structure output.
• Reasoning Traces: Utilizes "thinking traces" to integrate crystallographic and physical knowledge before generating coordinates, enhancing the scientific validity of the output.
• Reinforcement Learning (RL): Employs RL with verifiable rewards to improve the stability, validity, and property conditioning of the generated crystal structures.
• Pymatgen Integration: Provides a script to convert generated text descriptions directly into pymatgen.Structure objects for further analysis and use in materials science workflows.

Good For

• Researchers and developers in materials science and chemistry needing to generate crystal structures from high-level property descriptions.
• Automating the design and discovery of new materials with specific thermal expansion properties.
• Applications requiring the programmatic generation of valid and stable crystal structures with integrated physical constraints.