New materials are urgently needed to make better components used in sustainable energy. Technologies such as fusion and quantum computing need materials that can withstand high levels of radiation or support quantum computing, yet be safe, cost-effective, and sustainable. But these materials do not yet exist, and their discovery is a very difficult task, which includes the synthesis and testing of a large number of hypothetical materials.

“The materials found represent a very small fraction of hypothetical materials — like water droplets in the ocean,” Mingda Lee, a professor of nuclear science at MIT, wrote in an email.

The ability to perform tasks without human intervention makes the unmanned lab a “closed loop” system, which Polybot achieved last June.

One tool researchers are increasingly using to help with this discovery process is the unmanned lab, a lab system that combines advanced robotics with machine learning software to conduct experiments autonomously.

For example, Lawrence Berkeley National LaboratoryX Laboratory just opened last month and aims to find new materials that could help create better solar cells, fuel cells and thermoelectric technologies. (The lab says the “A” in its name is intentionally ambiguous, referring to autonomy, AI, abstraction, and acceleration in different ways.)

Another newly minted self-driving lab isnamed PolybotesV Argonne National Laboratory in Lemont, Illinois, runs a little longer than A-Lab. As a result, he climbed the ladder of laboratory autonomy to his own research in materials science. Polybot consists of chemical analysis equipment, computers with machine learning software and three robots. There is a synthetic robot that runs chemical reactions, a handler robot that cleans up reaction products, and a wheeled robot with a robotic arm that transports samples between stations. The robots are programmed using Python scripts and perform all the manual tasks in the experiments, such as sample loading and data collection.

The data collected during the experiments is then sent to a machine learning program for analysis. The software analyzes the results and suggests changes for the next series of experiments, such as adjusting the temperature, the amount of reagents, or the duration of the reactions. The ability to do all this without human intervention makes the unmanned lab a “closed loop” system, which Polybot achieved last June.

Argonne scientist Jie Xu, who began planning Polybot in 2019, said she wants the self-driving car lab to function as a resource that is “universally applicable and reconfigurable” for researchers of all stripes to take advantage of. Xu and his Argonne colleagues used Polybot to investigate electronic polymers, which are plastics that can conduct electricity. The hope is to create polymers that can make better and more sustainable versions of the technologies we use today, such as solar cells and biosensors.

Xu estimates that they will have to conduct half a million different experiments before they exhaust all possible ways to synthesize the target electronic polymer. An unmanned lab can’t try them all, not to mention human researchers, who can only generate about 10 molecules in two years, Xu said.

According to Xu, the unmanned labs help speed up the process of synthesizing new materials from two sides. One of them is the use of robotics: robots can perform synthesis and analysis of hypothetical materials faster than humans because robots can work continuously. Another way is to use machine learning to prioritize settings that are most likely to give the best result during the next experiment. Proper prioritization is important, Xu said, because the sheer number of experimental parameters, such as temperature and reagents, can be intimidating.

There are only a few self-driving car labs in the world today. However, this number will soon increase. Every national laboratory in the United States, for starters, is now building one.

Unmanned labs also have the advantage of generating large amounts of experimental data. This data is valuable because machine learning algorithms must be trained on a large amount of data in order to produce useful results. One laboratory is not capable of generating this amount of data on its own, so some laboratories have begun to combine their data with data from other researchers.

A-Lab LBNL also regularly contributes data to Project materials, which collects data from materials science researchers around the world. Milad Abolhasani, whose lab at North Carolina State University is studying self-driving cars, said increased sharing of open-source data is essential to the success of self-driving labs. But effective data exchange will require standardization of how data from laboratories are formatted and presented.

According to Abolhasani, there are only a few true unmanned laboratories in the world – laboratories that can work continuously without human intervention and without frequent breakdowns. That number could soon increase, he said, because every national lab in the United States is building one.

But there are still significant barriers to entry. Specialized robots and laboratory environments are expensive, and it takes years to build the necessary infrastructure and integrate robotic systems with existing laboratory equipment. Each time a new experiment is started, researchers may find that they need to make additional adjustments to the system.

Henry Chang, Xu’s colleague at Argonne, said they ultimately want Polybot’s machine learning capabilities to go beyond simple optimization experiments. He wants to use the system to make discoveries—create entirely new materials, such as polymers with new molecular structures.

A discovery is much more difficult to make because it requires machine learning algorithms to make decisions about where to go based on an almost unlimited number of starting points.

“For optimization, you can still define space, but for discovery, space is infinite,” Chan said. “Because you can have different structures, different compositions, different ways of processing.”

But the results in A-Lab show it’s possible. When the lab opened earlier this year, the researchers tried to synthesize completely new materials by running their machine learning algorithms on data from the Materials Project database. The self-driving car lab performed better than expected, showing promising results 70% of the time.

“We expected a success rate of around 30 percent at best,” wrote A-Lab principal investigator Gerd Seder.