AI Co-Scientists Cut 10,000 Experiments to 24, Transforming Lab Research

A laboratory at the Korea Institute of Science and Technology (KIST) in Seongbuk-gu, Seoul operates 24 hours a day without human presence. Inside the networked unmanned facility, robots and artificial intelligence work continuously. Human researchers need only enter a command on a computer screen requesting the development of a catalyst with specific properties. AI leads all experiments, including material synthesis, analysis, and redesign.

Four researchers led by Han Sang-soo, director of KIST's Computational Science Research Center, built this smart laboratory. Robots move around the lab synthesizing nanoparticles, but the robots that transfer reagents and operate equipment are merely tools. The core of the smart laboratory lies in the "AI co-scientist" autonomously deciding what experiments to conduct. When human researchers input desired properties, the AI co-scientist searches for candidate materials likely to possess those properties and calculates the conditions under which each candidate should be synthesized. If results fall short of expectations, the system adjusts conditions and repeats experiments, iterating toward optimal solutions.

Last year, KIST announced that AI in this smart laboratory could complete experiments that would require 10,000 trials by human researchers in just 200 attempts. Within a year, the laboratory's performance improved further. Recently, researchers succeeded in finding catalyst materials in only 24 experiments. Identifying the highest-performing catalyst took just over a week. The research team is now preparing patent applications focused on the material performance developed in the smart laboratory.

A wave of AI co-scientist adoption is sweeping through the global scientific community. This signals that AI is moving beyond simple experimental assistance or analytical tools to a stage where it designs experiments and participates in decision-making as a colleague to researchers. A prime example is the "Coscientist" system from Carnegie Mellon University. Researchers unveiled this system in the journal Nature in 2024. When instructed to conduct palladium catalyst reaction experiments, Coscientist independently searched papers and databases to learn existing knowledge and designed experimental protocols. It then directly commanded a robotic laboratory and succeeded in synthesis. When an error occurred in equipment control code during the experiment, the AI analyzed the equipment manual, diagnosed the error cause, corrected the code, and resumed the experiment. This is regarded as the first demonstration of AI performing everything from hypothesis generation to execution to error correction. Academia views this as "autonomy approaching that of a skilled research assistant or doctoral-level researcher."

Alexander V. Tobias and Adam Wahab of MITRE, a U.S. nonprofit research organization, recently published a paper categorizing the evolution of AI co-scientists in self-driving laboratories from Level 1 to Level 5. Level 1 represents the most basic automation stage, where robots simply repeat experiments designed by humans. Physical tasks such as reagent dispensing, mixing, and heating are central, with all judgment left to humans. Pharmaceutical quality control lines fall into this category. At Level 2, AI analyzes experimental results and presents patterns and correlations. Like DeepMind's AlphaFold, it excels at structure prediction or candidate identification, but actual experiment design and execution remain human domains. This is the "smart analyst" stage.

Many research institutions have currently reached Level 3. AI learns from experimental results and independently selects conditions for the next experiment. Using Bayesian optimization and active learning, it excludes experiments with low information value and explores only paths with high success probability. The KIST smart laboratory falls into this category.

The AI co-scientist stage currently drawing global attention is Level 4. At this stage, AI possesses autonomy beyond experiment design and execution to diagnose and correct problems occurring during experiments. AI recognizes errors and resolves them independently, with humans serving only as supervisors. Only a handful of examples exist worldwide.

Level 5 has not yet been realized. At Level 5, AI sets research objectives itself and determines research directions without human intervention. Reaching this stage would enable innovations where AI defines and explores uncharted territories in drug development, new materials discovery, and energy technology research. However, challenges also grow, including responsibility for research failures, potential generation of harmful substances, and ethical and safety regulations. Society must prepare more than technology at this stage.

AI co-scientists are expected to gain particular prominence in fields requiring repetitive experiments, including catalysts, batteries, displays, chemical materials, and drug discovery. However, these changes may pressure some experiment-focused research personnel toward job transitions or workforce reduction. Research communities anticipate a full-scale role transition from "researchers who conduct experiments" to "researchers who design and manage experimental systems."

In Korea, Vice Prime Minister and Minister of Science and ICT Bae Kyung-hoon recently announced a policy to actively promote AI co-scientist adoption with a 2026 target. If autonomous laboratory infrastructure development and AI-based research tool proliferation materialize, significant innovation in research speed and efficiency is expected. However, many obstacles remain, including investment in expensive automation equipment, improvements to equipment procurement review procedures, and training personnel in AI-experiment integration.