Chemical laboratories of the future will be able to work autonomously while also heeding human experience. A group led by Dr. Kwang-No An of the Korea Research Institute of Chemical Technology (KRICT) has developed an approach that allows scientists to explain their ideas in ordinary language, and artificial intelligence to use these hints to plan experiments.
The subject is so-called self-driving laboratories (SDL). Unlike conventional automated setups that simply repeat a given program, such systems choose for themselves which experiment to run next, relying on the results obtained. However, they have a weak spot: the "cold start problem." AI is forced to rediscover even well-known chemical regularities by conducting many costly experiments. In catalysis, where synthesis and analysis can take more than half a day per experiment, this is especially burdensome.
To get around the limitation, the team combined a large language model (LLM - an algorithm that understands and generates text) with Bayesian optimization (a method of searching for the best solution based on probabilistic models). The scientist formulates his knowledge and constraints in words, and the LLM turns them into structured prior information for the AI. In this way, expertise enters the system without complex programming or mathematical derivations.
In addition, the authors added an adaptive weight credibility detection (AWCD) mechanism. It monitors how well the hints agree with real data and automatically disables unreliable settings if evidence contradicting them appears. In tests with deliberately erroneous recommendations, conventional methods did not find the optimum, while the new system recognized the deception and restored search efficiency.
Testing on virtual environments built from real data showed a noticeable gain. For the Ugi reaction, the number of search experiments needed to reach target metrics fell from about 30 to fewer than 10. For P3HT-CNT conductive composites, from about 20 to three to five. According to the researchers' estimate, this is comparable to having about 40 known data points in advance and can save on the order of 20 days of initial screening in catalytic experiments where one run lasts at least half a day.
The team also released an interactive Co-Scientist interface that allows the direction of experiments to be adjusted in a natural-language dialogue. Further plans are to test the technology on the physical equipment of a self-driving laboratory. The results were published in the Journal of Chemical Information and Modeling.