With AI team member, Yale and Microsoft speed up battery innovation
Safe and scalable, redox flow batteries hold great potential as an alternative means of storing renewable energy on the grid. In terms of reducing cost and other factors, though, there’s still some work needed before they can be put to widespread use. Toward that end, Yale and Microsoft researchers are working with an AI system designed to consider the data like a scientist—a very speedy one at that—to figure out some crucial details.
Unlike conventional batteries, which usually store their energy in solid material, redox flow batteries store energy in liquid chemical solutions in tanks. Researchers have long searched for the optimal molecular compounds to use for the batteries. Among other qualities, they want components that are stable, highly soluble in water, and allow for high voltage. Identifying the best molecules for this is a process that requires building assumptions, updating them against evidence, and repeatedly calibrating what should be trusted, revised, or discarded over time.
To enhance and speed up that process, Yale’s David Kwabi and his research team used an approach developed by Microsoft called CLIO, cognitive loop via in-situ optimization. Kwabi’s lab is building aqueous organic redox flow battery (AORFBs), which are among the leading candidates for sustainable and long-duration energy storage. However, optimizing them for such qualities as stability, voltage, solubility, and efficiency is tricky. Plus, the molecule chosen for the battery needs to be relatively easy to synthesize in the lab.
According to Microsoft, CLIO “reasons by continuously reflecting on progress, generating hypotheses, and evaluating multiple discovery strategies.” That is, it’s designed to act like one more researcher on the team, capable of knowing when its first hypotheses are wrong, analyzing data from the lab, and correcting itself. In their study, the Yale-Microsoft research team noted that CLIO’s scientific reasoning is unique in that it reflects and calibrates its performance on the limitations of in-silico computational predictors.
“The big advantage of humans is that we can go into lab and do experiments and generate high-quality data,” said Kwabi, associate professor of chemical & environmental engineering. “The advantage of AI is that it can map large chemical design spaces and figure out which regions are interesting.”
As Kwabi explains, the goal is for the AI system to work with human experimentalists in a way that plays to the strengths of both.
“We have a scaffold around which we want to design a flow battery molecule,” he said. “The AI looks up a few interesting candidates. It then sends them to the humans. The humans test them, go back to the AI with the data, and then the AI says, ‘Okay, I'm going to form a hypothesis about what's happening.’”
In this case, CLIO initially suggested a molecular compound from the benzocinnoline family which was substituted with a benzylphosphonate group. While it performed well in some ways, the compound was inefficient at releasing the battery’s stored energy. Kwabi and his team provided the data to CLIO, which then suggested an altered version of the molecule. This time, the molecule was a success, which Kwabi said signals “a huge potential for making lots of progress in the future.”
"This work introduces a powerful new framework for advancing battery science with AI," he said. "By endowing an agent with the ability to reason from and adapt to experiments, we combine the strengths of human-led experimentation with AI's capacity to explore vast chemical design spaces—and we're only beginning to see what it can do."
More Details
Published Date
Jun 2, 2026
