A research team led by Professor Sung Beom Cho of the School of Advanced Materials Science and Engineering of Sungkyunkwan University (SKKU), in collaboration with the teams of Professors Jin Sung Park and Hyunsouk Cho of Ajou University and Professor Ju Li of the Massachusetts Institute of Technology (MIT), has developed a "closed-loop materials synthesis planning platform" that uses a large language model (LLM) to propose synthesis conditions and procedures for complex new materials and iteratively refine them based on experimental results.
The scientific community has recently been using various computational methods to rapidly identify promising new material candidates. However, actually synthesizing these materials requires extensive trial and error to determine specific conditions, such as which precursors to use and at what temperature and for how long the reaction should run. To reduce this trial and error, the research team devised a workflow in which the LLM searches existing literature for similar synthesis cases and proposes recipes suited to the new material.
The team first built a database by extracting key synthesis information — target materials, precursors, synthesis conditions sush as temperature, rpm, pressure, and time — from 4,407 open-access solid-state synthesis papers published in academic journals. The team then applied retrieval-augmented generation (RAG) method, which searches for similar existing synthesis cases based on a researcher's desired material and conditions and proposes candidate recipes accordingly. When the proposed recipes were compared against conditions reported in actual papers, they scored an average of around 4 out of 5 on key synthesis variables.
The team went on to use this AI platform to conduct synthesis experiments for an oxy-selenide-based solid electrolyte material for all-solid-state batteries, including previously unreported candidates. Under the 600°C condition initially proposed by the AI model, several impurity phases formed instead of the target material; but when the team fed these results back into the model, it proposed a follow-up recipe that progressively lowered the synthesis temperature. Based on this, the team sequentially tested conditions at 450°C and 400°C, and within just a few experiments succeeded in synthesizing a single-phase new material with no detectable impurities.
This research demonstrates the potential to reduce trial and error and shorten development timelines in materials design by combining human scientists' experimental experience with AI's ability to draw on a vast body of literature. Just as middle and high school students refer to existing recipes online and adjust ingredients and conditions based on their cooking results to develop recipes of their own, future scientists are expected to use AI to quickly incorporate experimental results, reduce unnecessary trial and error, and efficiently develop synthesis methods for a wide range of new materials. The findings were published in Advanced Materials, a leading international journal in materials science, and an image visualizing the research was selected as the back cover of the issue.