Comparison of FePt high-density magnetic recording media development between large language models and experimental experts: do LLMs recommend fluoride as a segregant?

We investigated segregant development for FePt high-density magnetic recording media using large language models (LLMs) and identified new fluoride-based materials. By increasing the temperature of the softmax function to select probable vocabularies, we successfully incorporated the prior knowledge involved in LLM training into the material selection process (exploratory generation) and identified LaF3 as an optimal material. Meanwhile, we independently conducted sputtering deposition of FePt-LaF3 nanogranular samples and verified whether the LLM could reproduce the results. The inhomogeneity in the surface chemical composition of FePt-LaF3 in the non-equilibrium state of sputtering were also reproduced by the LLM, leading to the identification of AlF3 as an alternative segregant. Structures and magnetic properties strongly dependent on the sputtering apparatus cannot be reproduced by simple LLM predictions, necessitating interactive data exchange between physical and cyber experiments. However, phenomena that can occur in physical experiments are almost accurately reproduced by LLM, demonstrating the usefulness of LLM predictions in material development. LLMs that generate responses from the most likely vocabulary tend to select well-known materials. To identify unknown materials as candidates, techniques such as exploratory generation or inputting unexpected results are required.