A team led by Asst. Prof. Chibueze Amanchukwu at the University of Chicago Pritzker School of Molecular Engineering has developed an AI model that identifies high-performing battery electrolytes using just 58 data points. The model, built through active learning, explored a virtual space of one million potential electrolytes and successfully pinpointed four new solvent candidates that match or exceed the performance of current state-of-the-art electrolytes. The breakthrough addresses a major bottleneck in materials discovery: traditional AI models require vast datasets—often millions of entries—to make reliable predictions. However, for emerging battery technologies, such data simply doesn’t exist yet. “Each experiment can take weeks or months to generate a single data point,” said Ritesh Kumar, co-first author and Schmidt AI in Science Postdoctoral Fellow at UChicago PME. “Waiting for millions of data points isn’t feasible when we’re racing against climate change and energy demands.” Instead of relying solely on computational simulations, the team integrated real-world experiments into the AI loop. The model would suggest a candidate electrolyte, the team would synthesize and test it in a working battery, measure its performance—particularly cycle life—and feed the actual experimental results back into the model. This “trust but verify” approach ensured that predictions were grounded in physical reality, not just theoretical calculations. Initially, the AI’s predictions carried high uncertainty due to limited training data. But through seven iterative cycles—each involving testing around 10 electrolytes—the model progressively refined its suggestions, narrowing down to four top performers. “We can’t eliminate the inefficiencies of AI completely,” Kumar said, “but we can use it where it excels: rapidly exploring vast chemical spaces that would be impossible to test manually.” The current model works by extrapolating from known molecules in existing databases. But the team envisions a future where AI becomes truly generative—capable of inventing entirely new molecular structures not found in any literature. “That would mean we’re no longer limited by what’s already been studied,” said co-first author Peiyuan Ma, Ph.D. “The model could propose molecules that have never existed before.” Future models will also need to evaluate electrolytes across multiple real-world criteria, not just cycle life. Safety, cost, stability, and energy density are all critical for commercialization. “We need AI to not just find good performers, but the best balanced performers,” Ma said. Ultimately, this approach helps scientists overcome human bias. Researchers naturally gravitate toward familiar chemical families, potentially overlooking revolutionary new materials. “AI can help us step outside our comfort zone,” Kumar said. “It gives us a way to explore uncharted chemical territory with confidence.”