Battery electrolytes aren\u2019t just one chemical, but a complex mixture of salts, solvents and additives interacting and reacting with each other.\u00a0<\/p>\n
Artificial intelligence has made great headway in helping select ideal materials to go into\u00a0that chemical soup. But a team from the University of Chicago Pritzker School of Molecular Engineering is using AI to generate the entire formulation, balancing the complicated tradeoffs and interactions that go into the electrolytes that make batteries possible.<\/p>\n
The research,\u00a0published in JACS Au,<\/a>\u00a0is the next step in the\u00a0Amanchukwu Lab\u2019s<\/a> ongoing development of an AI for battery work, which they\u2019ve nicknamed \u201cElectrolyteGPT<\/a>.\u201d<\/p>\n \u201cNext-generation battery electrolytes must meet multiple, often conflicting property requirements,\u201d said Jaemin Kim, SM\u201925, the first author on the study. \u201cWith the model\u2019s capability of generating outputs under diverse conditions, ElectrolyteGPT is able to generate novel candidates satisfying the desired properties simultaneously.\u201d<\/p>\n The AI doesn\u2019t 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.\u00a0<\/p>\n Synthesizing the AI\u2019s recommendations found several novel compositions that performed as well as top-of-the-line electrolytes in lithium metal batteries. It\u2019s an important step toward the long-term goal of finding electrolytes that outperform the current best, said corresponding author Neubauer Family Asst. Prof. Chibueze Amanchukwu<\/a>.<\/p>\n \u201cWe had a number of compositions that performed on par with the state of the art, and so that was exciting for us,\u201d Amanchukwu said. \u201cWe can generate compositions that can mimic what some of the best scientists have done, but there\u2019s still lots of work ahead.\u201d<\/p>\n Exploring a vast chemical space<\/p>\n Many estimate the number of potential molecules for battery electrolytes is 10 to the 60th power\u2014more 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 the human lifespan.<\/p>\n That\u2019s just the molecules themselves, not the practically infinite possible ways to combine them in different formulations.<\/p>\n \u201cWhile it\u2019s infeasible to explore the entire near-infinite electrolyte space, generative AI can navigate the \u2018unmapped\u2019 areas of chemistry and generate a molecule that has never been synthesized before,\u201d Kim said.<\/p>\n The AI generates theoretical molecules at a rate human researchers could never match, pulling ones it \u201cthinks,\u201d 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.\u00a0<\/p>\n AI is often used for drug discovery, which set up an early hurdle for Amanchukwu\u2019s team. Most existing GPT models were created to find molecules that make good drugs, not good batteries.\u00a0<\/p>\n \u201cIf you use what is available in the literature, it will generate drug-like molecules. That\u2019s not relevant for us,\u201d Amanchukwu said. \u201cWe curated a data set that has electrolyte-relevant compounds so that the GPT model only knows about electrolytes. Then if you say, \u2018Generate new solvent molecules,\u2019 it generates compounds that look like they could be electrolytes.\u201d<\/p>\n Next, they had to train the AI to generate only electrolyte materials that hit certain parameters. There\u2019s no point in training an AI to create low-performing batteries, so they set standards for ionic conductivity, oxidative stability, Coulombic efficiency and viscosity.<\/p>\n 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.\u00a0<\/p>\n Inventing the fLine<\/p>\n Line notations are a way of describing complex chemical structures using language a computer understands. For example, calling sodium chloride \u201csalt\u201d would get the point across to a person but a machine would be confused by all the other possible salts in the world.\u00a0<\/p>\n 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.\u00a0<\/p>\n 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.<\/p>\n This allows an AI to understand the entirety of an electrolyte, not only the chemicals that go into it.<\/p>\n \u201cThat is useful for not just electrolytes. It is useful for mixtures in general,\u201d Amanchukwu said. \u201cNow you can actually generate a complete electrolyte formulation with multiple different salts, multiple different solvents at different concentrations and at different mixture ratios.\u201d<\/p>\n Amanchukwu said this is an important step toward the ultimate goal: truly generative electrolyte AI.\u00a0<\/p>\n \u201cRight 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\u2019s theoretical suggestions in the real world,\u201d he said. \u201cWe are interested in seeing if we can make these models bigger and better.\u201d<\/p>\n Citation: “Generative Electrolyte Solvent and Formulation Discovery,” Kim et al,\u00a0JACS Au, April 9, 2026. DOI:\u00a010.1021\/jacsau.5c01628<\/a><\/p>\n