AI research partner cuts new-material development time by up to 75%

A researcher says, "Raising the temperature should improve the yield" — and an AI instantly understands and designs the next experimental conditions on its own. Korean scientists have developed technology that does exactly that, transferring a researcher's experience and knowledge directly to AI to eliminate unnecessary experiments and dramatically cut the time and cost of developing new materials and catalysts.

The Korea Research Institute of Chemical Technology announced Wednesday that a research team led by Ahn Kwang-no of its Chemical Platform Research Division had developed "human-in-the-loop optimization" technology, in which AI interprets a researcher's natural-language knowledge and incorporates it into experimental planning.

At the heart of the technology is a large language model (LLM) that understands experience and knowledge conveyed by a researcher in speech or writing, then passes that information to a separate AI that determines experimental conditions. Researchers and AI can collaborate to design experiments through conversation, with no complex programming or mathematical formulas required.

The scientific community has been paying growing attention to "self-driving laboratories" (SDLs) — in which AI learns from experimental results and independently decides on the next experiment — as a next-generation approach to research and development.

Where conventional automated equipment simply repeats a fixed set of experiments, an SDL has AI analyze results and search for optimal outcomes by adjusting conditions such as temperature, pressure and raw-material ratios.

Existing SDL systems, however, could not draw on knowledge researchers already possessed and had to repeat experiments from the outset — a significant limitation.

This was particularly costly in catalyst and new-material development, where a single synthesis and analysis can take half a day or more, making unnecessary experiments a major driver of longer research timelines and higher costs.

To address this, the team combined an LLM capable of understanding researchers' language with a "Bayesian optimization AI" that finds optimal conditions with fewer experiments.

When a researcher shares experience — for example, "increasing temperature raises yield" — the LLM converts that input into prior information and constraints needed for the experiment. The optimization AI then uses that information to select the next set of experimental conditions.

The system also incorporates a safeguard: if a researcher's experience or judgment turns out to be incorrect, the AI filters it out on its own.

The team's "adaptive weight credibility detection" (AWCD) technology continuously checks how closely a researcher's advice aligns with actual experimental data. When advice diverges from the data beyond a set threshold, the system automatically discards that input and switches to a search based solely on real data.

In virtual experiments using actual chemical-reaction and materials data, the number of searches needed to find optimal conditions for reactions combining multiple raw materials fell from about 30 to fewer than 10.

For electrically conductive composite materials, the number of searches dropped from roughly 20 to just 3 to 5 — reducing the experiment count to less than one-quarter of what was previously required.

According to the research team, this efficiency gain is equivalent to having about 40 real experimental data points on hand from the start. Applied to catalyst experiments that each take half a day or more, the technology could shorten the initial search phase by roughly 20 days.

The team also developed and released an "AI co-scientist" interface that lets researchers converse with the AI mid-experiment to adjust the research direction in real time.

"This technology is core software that acts as an experimental partner — and the brain of an autonomous laboratory — helping researchers and AI support each other and conduct research together," Ahn said. "It will contribute to significantly reducing the time and cost invested in research and development."

The findings were published in the August issue of the Journal of Chemical Information and Modeling.