Plain-language advice cuts AI lab searches for catalysts, in simulation

A chemist types a plain sentence into the system: “Increasing the temperature improves yield.” Within the same session, the AI has already changed the order in which it tests conditions. That is the core of a new framework from the Korea Research Institute of Chemical Technology (KRICT), described by Dr. Gwang-Noh Ahn's team as a way to let scientists hand their know-how directly to a self-driving laboratory.

The mechanism works in three steps. A large language model (LLM), the kind of program that reads and writes everyday text, converts the researcher's remark into structured prior information and constraints. A Bayesian optimization (BO) algorithm, a method that weighs past results to choose the most promising next experiment, uses that information to pick the next run. The results go back into the loop, and the researcher can revise the guidance in natural language at any time. The team calls the whole approach human-in-the-loop optimization (HILO). Conventional automated equipment repeats predefined procedures; here the AI decides which experiment comes next.

The risk is obvious: a wrong hunch can send the AI in the wrong direction. To address it, the team added a safeguard it calls adaptive weight credibility detection (AWCD). It compares the advice with the experimental data and, when enough contradictory evidence accumulates, switches off the misleading prior and returns to purely data-driven search. In tests with deliberately incorrect guidance, conventional methods failed to find the optimal conditions, while the new framework detected the error and recovered its performance. When the advice is accurate, AWCD keeps it active.

The gains reported are from about 30 to fewer than 10 searches for the Ugi reaction, and from about 20 to three to five searches for the P3HT-CNT composite. The team estimates the effect is comparable to having about 40 experimental data points available in advance, which could save around 20 days of initial exploration in catalyst experiments that take half a day or more per run. These are virtual experiments built on real data, not runs on a physical robot, and the team says it still has to validate the system on physical self-driving lab equipment. No date for that work is given.

For a research group, the immediate benefit would be fewer wasted runs in expensive fields such as catalysts and new materials, where the team expects the approach to cut development time and costs. Because the guidance is written in ordinary language, a chemist without programming skills could steer the search, and an interactive interface called Co-Scientist, which the team has released, allows changes in direction during an optimization. Whether that holds up on real equipment, and on problems far from the two tested reactions, is what the next experiments will decide.