The Pulse
Microsoft Says AI Designed a Vanadium-Free Flow Battery Candidate
Microsoft says its Discovery Engine helped Yale researchers design and validate a novel organic molecule for long-duration redox-flow batteries. The work uses AI agents, laboratory measurements and stored negative results to refine battery
AI.info Team ·
Microsoft is presenting a battery experiment as evidence that AI can do more than generate chemical candidates: it can preserve failed experiments, revise its strategy and help researchers reach a testable result.
In a post published on Microsoft’s Tech Community blog on September 7, 2026, the company said its Discovery Engine helped design a novel organic negolyte for aqueous organic redox-flow batteries. The candidate, developed with Yale Engineering, Canam Bioresearch and Pacific Northwest National Laboratory, is intended to reduce reliance on vanadium, a costly mineral used in many grid-scale flow batteries.
Microsoft describes the result as its first end-to-end validated scientific outcome from Discovery, combining computational predictions with laboratory measurements. The company has not published a full performance dataset showing that the candidate is ready for commercial deployment, so the announcement represents a research result rather than a finished battery product.
Three design rounds, not one lucky prediction
The work began with the benzo[c]cinnoline chemical scaffold, which Microsoft and its partners modified in search of an organic molecule suitable for flow-battery electrolytes. Discovery Engine used CLIO, short for Cognitive Loop via In-Situ Optimization, to run multiple lines of reasoning and adjust its approach as new evidence arrived.
Microsoft says the system organized the campaign into three rounds. The first two rounds explored a broad range of molecular changes, but many candidates failed to meet the required chemical or electrochemical conditions. In the second round, none of the parallel explorations produced a candidate suitable for advancement.
Those failures became part of the input to the third round through Discovery Bookshelf, a knowledge system that records links between hypotheses, predictions, measured outcomes and confidence levels. Microsoft says four of six explorations in the third round succeeded after the system used the earlier results to change how it interpreted its computational tools.
AI learned to distrust an imperfect predictor
The most revealing adjustment involved a redox-potential predictor. Microsoft says the model did not produce numerically accurate values when compared with laboratory measurements, but its rankings still pointed researchers toward relatively better candidates.
Discovery therefore stopped treating the predictor as a source of precise measurements and used it as a directional ranking tool. That distinction allowed the system to retain useful information from an imperfect model instead of discarding the model altogether.
Microsoft says this process also helped the agents avoid repeating unsuccessful molecular modifications. The company frames that ability as a form of scientific memory: a record of why previous approaches failed, rather than a database containing only successful results.
This work introduces a powerful new framework for advancing battery science with AI. 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.
David Kwabi, associate professor of chemical and environmental engineering, Yale University
Why flow batteries need new chemistry
Redox-flow batteries store energy in liquid electrolytes held in external tanks. Pumps move the liquids through an electrochemical cell during charging and discharging, allowing the storage system’s energy capacity to increase by adding larger tanks rather than building more battery cells.
Vanadium-based systems have attracted attention for grid storage because the same element can operate on both sides of the battery, reducing certain forms of cross-contamination. Vanadium prices, supply constraints and the cost of large installations have encouraged researchers to investigate organic alternatives that can be made from more widely available materials.
Organic molecules introduce their own problems. A useful negolyte must combine the right redox potential with high water solubility, chemical stability, reversible electrochemical behavior and a synthesis route that works outside a computer simulation. Improving one property can damage another, which makes candidate selection a multi-variable problem.
What Microsoft has actually demonstrated
Microsoft’s announcement establishes that Discovery Engine supported a closed-loop campaign in which AI-generated designs were tested against laboratory results and then revised. Yale researchers led the experimental characterization, while Microsoft says the system coordinated literature analysis, candidate generation, computational screening and interpretation of results.
The announcement does not establish a commercial energy density, cycle life, cost per kilowatt-hour or full-scale prototype. It also does not show that the organic candidate outperforms vanadium systems in an operating grid installation.
That distinction matters because battery discovery has two separate hurdles: finding a molecule with promising properties and proving that it survives repeated operation in a practical device. Microsoft’s earlier collaboration with PNNL followed a similar path when AI screened 32.6 million candidate materials and helped identify a solid-state electrolyte using approximately 70% less lithium than conventional lithium-ion batteries. That material was synthesized and tested in a working prototype, but Microsoft described further validation and optimization as ongoing.
The value may be in preserving failure
Scientific projects often record successful measurements more carefully than failed directions. Microsoft argues that the discarded paths can be just as useful when an AI system must decide what to test next, particularly when computational models disagree with experiments.
Discovery Bookshelf is designed to retain those relationships across a campaign. Instead of storing only candidate structures, it records the reasoning behind a choice, the evidence supporting it, the result that followed and the confidence assigned to each conclusion.
For the battery project, the practical payoff was not an autonomous laboratory producing a finished cell. It was a narrower result: an AI system helped researchers convert unsuccessful experiments into a better next round, leading to a novel organic candidate that was validated through in-lab measurements.