Artificial intelligence is increasingly helping scientists: analyzing data, predicting the properties of substances and searching for promising materials. But most such systems exist only on a screen - they cannot conduct an experiment in a laboratory themselves; that requires a human or separate automation.

Researchers from University College London (UCL), the Hong Kong University of Science and Technology (Guangzhou) and other organizations have developed a system that combines digital intelligence with a robot. They called it embodied science agent (ESA). The paper was published on the arXiv preprint server, meaning it has not yet undergone journal peer review.

The new agent consists of three parts. The digital part, based on accumulated data and records of past experiments, decides what to investigate next. To choose conditions it uses Bayesian optimization - a mathematical method that helps find the best options among a huge number of possibilities, relying on the results of previous attempts. Then the embodied part - the robot - performs the necessary actions, adapting to the position of the instruments. The third component - the interface - links the plan and its physical execution.

The authors tested the system on wastewater treatment to remove dye. The agent had to choose conditions from more than 200 million combinations, with the results of each experiment influencing the next. ESA ran 15 cycles in a row. During that time the clarity of the solution rose from 82% to 99.7%, and the residual dye concentration fell by 98.5% - almost complete decolorization. In a separate test the robot handled the work even when the electrode stand was moved to different positions.

It is important to note: what was measured was decolorization. This does not prove that the dye was completely decomposed or that other pollutants were removed from the water. In addition, these are laboratory tests, not a ready industrial installation.

Nevertheless, the work shows that it is possible to create more adaptive scientific agents capable of both planning and conducting experiments. In the future the authors propose improving the system and testing it on a broader range of experiments from different fields of science.

Source: Tech Xplore