Researchers develop conversational AI for self-driving labs

by National Research Council of Science and Technology

edited by Swati Mestri, reviewed by Andrew Zinin

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

Add to Preferred Sources When researchers share their experience in everyday language, such as "Increasing the temperature improves yield," a large language model (LLM) translates their input into structured prior information that AI can use. The Bayesian optimization AI then uses this knowledge to select the next experiment. Experimental results are fed back to the AI for further learning, while researchers receive updates and can revise their guidance in natural language at any time.

An adaptive weight credibility detection (AWCD) mechanism also monitors whether the guidance aligns with experimental data and automatically disables misleading prior information when sufficient contradictory evidence emerges. Credit: Korea Research Institute of Chemical Technology (KRICT) Korean researchers have developed an AI technology for self-driving laboratories (SDLs) that incorporates researchers' experience and knowledge, expressed in natural language, into autonomous experimental planning.

A research team led by Dr. Gwang-Noh Ahn at the Korea Research Institute of Chemical Technology (KRICT) developed a human-in-the-loop optimization (HILO) framework that uses large language models (LLMs) to translate researchers' knowledge into information for experimental optimization.

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

The technology is expected to reduce costly trial-and-error experiments and accelerate the development of new materials and catalysts.

Self-driving laboratories autonomously identify optimal experimental conditions by learning from previous results. Unlike conventional automated equipment that repeats predefined procedures, SDLs independently select subsequent experiments.

However, conventional SDLs often face a cold-start problem, requiring AI to rediscover even well-established chemical knowledge through repeated experiments. This is particularly costly in catalyst research, where synthesis and analysis can take more than half a day per experiment.

To overcome this limitation, the team integrated an LLM with Bayesian optimization (BO). The LLM converts researchers' natural-language guidance into structured prior information and constraints, allowing AI to incorporate expert knowledge without requiring complex programming or mathematical modeling.

The team also introduced adaptive weight credibility detection (AWCD) to prevent misleading expert knowledge from compromising optimization. AWCD monitors consistency between prior knowledge and experimental data, automatically disabling unreliable priors and reverting to data-driven optimization when sufficient contradictory evidence emerges.

In tests involving deliberately incorrect guidance, conventional methods failed to identify optimal conditions, while the new framework detected misleading information and recovered optimization performance.

Validation using virtual experimental environments based on real data demonstrated substantial improvements. For the Ugi multicomponent reaction, the number of searches required to achieve target performance decreased from approximately 30 to fewer than 10. For electrically conductive P3HT-CNT composites, searches dropped from approximately 20 to just three to five.

These results indicate an effect comparable to having approximately 40 experimental data points available in advance, potentially saving around 20 days of initial exploration in catalyst experiments requiring half a day or more per run.

The team also developed and released an interactive Co-Scientist interface, enabling researchers to adjust experimental directions through natural-language dialogue during optimization. The researchers plan to validate the technology using physical self-driving laboratory equipment.

Ahn of KRICT said, "This technology serves as both a research partner that enables scientists and AI to collaborate and a core software platform functioning as the brain of a self-driving laboratory. We expect it to significantly reduce research and development time and costs."

Amirreza Mottafegh et al., Adaptive human-in-the-loop optimization using language-guided priors for chemical experiments, Journal of Chemical Information and Modeling (2026). DOI: 10.1021/acs.jcim.6c00976 Journal information: Journal of Chemical Information and Modeling

Swati Mestri holds a bachelor's degree in Electronics Engineering and has worked as a content editor since 2019. She has experience editing research documents across technology, health care, and materials science, and has a particular interest in technology and space. Full profile →

Master's in physics with research experience. Long-time science news enthusiast. Plays key role in Science X's editorial success. Full profile →