Using the open source base models like Llama-8B, Qwen-14B and Mistral-7B models, I am fine-tuning Large Language Models to interpret chemical compositions, structure and predict the key magnetic properties like Curie temperature, Magnetization, Coercivity, Magnetocrystalline Anisotropy. This involves creating custom tokenizers for chemical formulas and utilizing the Unsloth library for efficient training on consumer GPUs. We are also working on inverse design of functional magnetic materials using LLMs
Applying Proximal Policy Optimization (PPO) and GRPO to optimize material structures for specific magnetic traits.
Using the api based smart LLMs like GPT5 and Gemini models, we are collecting the experimental based scientific knowledge from the peer reviewed scitific articles.