AI generates entire chemical formulations for battery electrolytes

Battery electrolytes aren’t just one chemical, but a complex mixture of salts, solvents and additives interacting and reacting with each other. 

Artificial intelligence has made great headway in helping select ideal materials to go into that chemical soup. But a team from the University of Chicago Pritzker School of Molecular Engineering (UChicago PME) is using AI to generate the entire formulation, balancing the complicated tradeoffs and interactions that go into the electrolytes that make batteries possible.

The research was published in JACS Au. It is the next step in the Amanchukwu Lab’s ongoing development of an AI for battery work, ElectrolyteGPT.

“Next-generation battery electrolytes must meet multiple, often conflicting property requirements,” said first author Jaemin Kim, SM’25. “With the model’s capability of generating outputs under diverse conditions, ElectrolyteGPT is able to generate novel candidates satisfying the desired properties simultaneously.”

The AI doesn’t just set the ingredients, but the concentrations, mixture ratios and other facets of the blend, hitting targets the researchers set on everything from conductivity to stability to viscosity. 

Synthesizing the AI’s recommendations found several novel compositions that performed as well as top-of-the-line electrolytes in lithium metal batteries. It’s an important step toward the long-term goal, finding electrolytes that outperform the current best, said corresponding author Neubauer Family Asst. Prof. Chibueze Amanchukwu.

“We had a number of compositions that performed on par with the state of the art, and so that was exciting for us,” Amanchukwu said. “We can generate compositions that can mimic what some of the best scientists have done, but there’s still lots of work ahead.”

Exploring a vast chemical space

Many estimate the number of potential molecules for battery electrolytes is 10 to the 60th power, more than all the stars in the sky. Exploring each of those molecules for battery components, cancer drugs or other previously undreamed materials is simply beyond human lifespan.

That’s just the molecules themselves, not the practically infinite possible ways to combine them in different formulations.

“While it’s infeasible to explore the entire near-infinite electrolyte space, generative AI can navigate the ‘unmapped’ areas of chemistry and generate a molecule that has never been synthesized before,” Kim said.

The AI generates theoretical molecules at a rate human researchers could never match, pulling ones it “thinks,” based on training data, would be good for a particular purpose. People then lab-test the materials the AI found the same they would test a material a researcher suggested. 

AI is often used for drug discovery, which set up an early hurdle for Amanchukwu’s team. Most existing GPT models were created to find molecules that make good drugs, not good batteries. 

“If you use what is available in the literature, it will generate drug-like molecules. That’s not relevant for us,” Amanchukwu said. “We curated a data set that has electrolyte-relevant compounds so that the GPT model only knows about electrolytes. Then if you say, ‘Generate new solvent molecules,’ it generates compounds that look like they could be electrolytes.”

Next, they had to train the AI to generate only electrolyte materials that hit certain parameters. There’s no point in training an AI to create low-performing batteries, so they set standards for ionic conductivity, oxidative stability, Coulombic efficiency and viscosity.

So far, this is all becoming established practice in the cutting-edge field of AI for materials discovery. The innovation of this new research was inventing a new line notation called the fLine. 

Inventing the fLine

Line notations are a way of describing complex chemical structures using language a computer understands. For example, calling sodium chloride “salt” would get the point across to a person but a machine would be confused by all the other possible salts in the world. 

In SMILES, one of the most common chemical languages, sodium chloride would be [Na+].[Cl-], a few keystrokes that transmit a massive amount of information. 

Based on SMILES, fLine is a new language the team developed that describes not only structure, but also includes notations for solvent ratio, salt concentrations, temperature and the other moving parts that go into a mixture. It could also be adapted to include variables such as current density and capacity if needed as well as other chemical languages beyond SMILES.

This allows an AI to understand the entirety of an electrolyte, not only the chemicals that go into it.

“That is useful for not just electrolytes. It is useful for mixtures in general,” Amanchukwu said. “Now you can actually generate a complete electrolyte formulation with multiple different salts, multiple different solvents at different concentrations and at different mixture ratios.”

Amanchukwu said this is an important step toward the ultimate goal: truly generative electrolyte AI. 

“Right now, even with the limited data as well as the limited parameters that we run, we can actually generate compositions that we experiment in. We can verify the AI’s theoretical suggestions in the real world,” he said. “We are interested in seeing if we can make these models bigger and better.”

Citation: "Generative Electrolyte Solvent and Formulation Discovery," Kim et al, JACS Au, April 9, 2026. DOI: 10.1021/jacsau.5c01628

---

---