What if an AI could screen battery materials from their chemical formula alone, before a single crystal structure had been worked out?
That is the premise of IonNet, a framework built at Cornell University's AI for Science Institute by Fengqi You and postdoc Zhilong Wang. The results appeared in Science Advances on August 7. Electrolytes shuttle ions between electrodes, so they set a battery's energy density, charging speed, and lifetime. Minor shifts in composition change cell performance dramatically. Most AI models for materials need a reliable crystal structure, and new compounds rarely have one.
IonNet predicts lithium-ion mobility from composition alone. Across about 4,500 stable compounds, it flagged 87 fast-ion conductors (FICs). From roughly 5 million substituted compositions it surfaced nearly 63,000 more candidates. The authors then ran physics-based simulations on 20 of the most promising picks. Thirteen held up as fast-ion conductors.
Simulations are the bottleneck. One material can consume tens of thousands of CPU hours. "IonNet is a fast front-end that prioritizes candidates before committing major experimental or computational resources," You said.
The screening is useful. The design rules may matter more. IonNet found that light magnesium doping, around 1%, speeds up lithium-ion movement. Heavy doping, around 10%, blocks it by crowding migration paths and distorting the lattice. That kind of qualitative insight is exactly what trial-and-error development buries.
The stakes justify the effort. Global lithium-ion demand is projected to hit 2,000 gigawatt-hours (GWh) by 2030. Solid-state batteries look like the core of next-generation storage. Thermal-runaway accidents already cost the industry more than $1 billion a year. Replacing flammable liquid electrolytes with solids is a safety argument and a capacity argument at once.
Scale is the open question.
None of the 87 candidates is a factory part yet. A confirmed fast-ion conductor still needs a working pairing with cathode and anode, a manufacturing process that stays repeatable, and years of qualification before volume shipments. The AI shortens the shortlist, not the pipeline. Announced capacity and commissioned capacity remain different numbers, and that gap is where battery hype tends to die.
The economics argue for faster screening. Battery makers already spend heavily on electrolyte R&D, and most of that spend lands on compounds that never reach a cell. An AI front-end that pares candidates down before simulation does not remove the manual work, it redirects it. Ten thousand hours of compute spent on a confirmed fast-ion conductor beats spreading the same budget across dead ends.
Cornell also published a companion study in Nature Communications on June 5, attacking the problem from the other end. Those batteries generate voltage from differences in electrolyte composition rather than different electrode materials. Both projects sit under one umbrella, an attempt to turn electrolyte discovery from a manual search into a guided one.
IonNet changes how the field picks candidates, not the lab work itself. You framed the shared goal plainly: "to move battery design from trial and error toward rational design guided by AI, chemistry and physical insight." Candidates screened against physical rules first are cheaper by orders of magnitude than whatever the next guess would have cost.
