In 2016, Samsung's Galaxy Note 7 caught fire around the world, leading to a recall of roughly 2.5 million units. The cause was the organic solvent-based electrolyte inside the lithium-ion battery. While this liquid excels at transporting lithium ions rapidly, it is also highly flammable. As electric vehicles and grid-scale energy storage systems push toward larger capacities, this contradiction only grows more serious.
All-solid-state batteries, which replace the liquid with a solid electrolyte, solve this problem in principle. Using a non-flammable solid material dramatically reduces the risk of fire. Combined with lithium-metal anodes, they also promise higher energy density.
However, moving lithium ions fast enough through a solid has remained the field's biggest obstacle.
A 2011 Discovery Revealed the Possibility—and Left a Gap Beyond It
In 2011, a group led by Professor Ryoji Kanno at Tokyo Institute of Technology announced the sulfide-based solid electrolyte (LGPS). Its room-temperature ionic conductivity was 12 mS/cm—a figure that exceeded that of organic liquid electrolytes, and the first result to demonstrate that solid electrolytes could match liquids in performance.
The search for solid electrolytes intensified after this discovery. But just as the discovery of LGPS itself owed much to painstaking exploration of ternary phase diagrams and a chance observation (that conductivity did not drop under certain polishing conditions), the search for new materials has continued to depend heavily on researchers' intuition and effort.
Screening via first-principles calculations (density functional theory, DFT) or molecular dynamics (MD) simulations has also been attempted, but the computational cost is enormous. Evaluating the ionic conductivity of a single material through physical simulation can require tens of thousands of CPU hours. Even with high-throughput DFT screening and preliminary narrowing, the number of candidates that can realistically be handled is said to top out at a few hundred.
Meanwhile, recent machine learning approaches have offered a way to break through this wall. However, most existing AI models require accurate crystal structure data as input. For new materials or compounds only recently reported experimentally, the crystal structure is often undetermined. This constraint—the inability to screen candidates lacking structural data—has been a bottleneck for AI adoption in this field.
IonNet's Design Philosophy: Predicting Ion Movement from Composition Alone
IonNet, a framework presented on August 7, 2026 in Science Advances by Professor Fengqi You (Chemical and Biomolecular Engineering) and postdoctoral researcher Zhilong Wang at Cornell University, removes this constraint head-on.
IonNet's full designation is reflected in the paper's title: a "descriptor-guided transfer learning framework." As the name suggests, the only input required is chemical composition—not crystal structure.
As Professor You explained in a Cornell University press release: "Many AI models for materials require a reliable crystal structure, but that's often not available for new or experimentally reported compounds. IonNet predicts ion mobility from chemical composition even without an accurate crystal structure."
This design changes the "order of operations" in the materials discovery workflow. Traditionally, researchers would first determine the crystal structure, then evaluate it using AI or simulation based on that structure. With IonNet, screening can proceed immediately once the composition formula is known. Determining the crystal structure can wait until after candidates have been narrowed down.
From 4,500 to 5 Million: What Expanding the Search Space Revealed
The research team applied IonNet in two stages.
In the first stage, they screened roughly 4,500 known stable compounds and identified 87 candidate fast ion conductors. In the second stage, they systematically substituted elements into existing compositions, generating approximately 5 million compositions, and ran them through IonNet. As a result, about 63,000 were judged to have properties promising for fast ion conductors.
This figure of 63,000 stands out when compared to the scale of searches possible with conventional methods. Given that high-throughput DFT screening can typically handle only a few hundred candidates, IonNet has expanded the search space by more than two orders of magnitude.
| Item | Conventional Methods (DFT/MD) | IonNet |
|---|---|---|
| Input data | Crystal structure + composition | Composition only |
| Computational cost per material | Tens of thousands of CPU hours | Substantially lower (screening stage) |
| Number of candidates that can be screened | A few hundred or so | ~5 million screenable |
| Number of candidates identified in this study | (Tens in prior studies) | ~63,000 |
What a 65% Hit Rate Shows—and What It Doesn't
To verify whether the AI's predictions could be trusted, the research team selected about 20 candidates and re-verified them using physics-based simulation. Of these, 13 were confirmed as fast ion conductors—a hit rate of 65%.
This figure requires careful interpretation. The 65% does not mean "the AI's predictions were entirely correct." Rather, it reflects a staged verification process: among the 20 candidates selected for costly simulation, 13 turned out to be hits. Whether the remaining 7 misses stemmed from IonNet overestimating their potential, or from differences in simulation conditions, would require examining the paper's details further.
Moreover, confirmation via physical simulation remains a computational verification—the materials have not actually been synthesized and measured. Engineering-critical properties such as chemical stability of the candidate materials, reactions at electrode interfaces, and long-term cycling characteristics remain unevaluated.
Professor You has been clear about IonNet's positioning: "IonNet is not meant to replace experiments or high-fidelity simulations. It's a fast front end for prioritizing candidates before committing large experimental or computational resources."
AI That Can Explain "Why It's Fast"
Another function of IonNet is extracting chemical design rules. Rather than simply producing a list saying "this composition looks promising," it can identify which combinations of elements or chemical environments are associated with fast lithium-ion movement.
As Professor You put it: "We didn't just want a list of promising materials. We wanted to understand the chemical principles behind why lithium ions move fast—because that's how AI can actually inform experimental design."
This "explainability" is a factor that determines success or failure for AI applications in materials science. A black-box model, even with high predictive accuracy, offers experimentalists no guidance on what to synthesize next. The design rules IonNet presents aim to be in a form that synthetic chemists can intuitively understand and translate into experimental plans.
A Second Study That Broke Through the "0.059 V Wall" of Concentration Cells
The same research team published another result on June 5, 2026 in Nature Communications. This work concerns not solid electrolytes but the voltage of concentration batteries.
A concentration battery uses the same redox couple at both electrodes, generating electromotive force purely from a concentration difference in the electrolyte. According to the Nernst equation, formulated by Walther Nernst in 1889, at room temperature a tenfold difference in concentration yields an electromotive force of only about 0.059 V. Because of this thermodynamic constraint, concentration batteries have not been regarded as a practical means of energy storage.
A joint research team from Cornell University and the University of Puerto Rico, Río Piedras circumvented this constraint by manipulating how ions in the electrolyte coordinate with surrounding molecules (solvation structure). On the cathode side, a high-concentration electrolyte raises the redox potential of the positive electrode; on the anode side, a strongly coordinating electrolyte lowers the free-ion concentration, thereby lowering the redox potential of the negative electrode. This asymmetric electrolyte design achieved voltages of 0.6–0.7 V in zinc-based and copper-based concentration cells—more than ten times the conventionally assumed value of 0.059 V.
Furthermore, by combining this electrolyte strategy with ordinary electrode materials, the team achieved an operating voltage of 2.2–2.5 V in a full aqueous cell. Given that an alkaline AA battery has a nominal voltage of 1.5 V, this means an aqueous battery surpassed that voltage.
| Item | Conventional Concentration Cells | This Study's Results |
|---|---|---|
| Voltage (concentration difference alone) | ≤0.059 V (per tenfold concentration difference) | 0.6–0.7 V |
| Voltage multiplier | Baseline | ~12x conventional expectation |
| Full-cell voltage combined with ordinary electrodes | Not applicable | 2.2–2.5 V |
| Comparison (alkaline AA battery) | 1.5 V | 2.2–2.5 V |
From Trial and Error to Rational Design—How Far There Is Still to Go
What the two studies share is a common goal: shifting battery design from a cycle of "build and measure" to rational design guided by AI and physicochemical insight. As Professor You put it: "These electrolyte studies share the same big-picture goal—moving battery design away from trial and error and toward rational design guided by AI, chemistry, and physical insight."
That said, all 63,000 candidates identified by IonNet remain computational predictions. Issues directly relevant to practical application—synthesizability of the candidate materials, stability against air and moisture, side reactions occurring at electrode interfaces, and resistance to volume changes during charge-discharge cycling—remain unverified. Even the 65% hit rate is based on cross-checking against physical simulation; verification through laboratory synthesis and measurement has not yet occurred.
As for the concentration battery results, whether a voltage of 0.7 V can be reconciled with practical energy density and cycle life awaits future evaluation. It also remains uncertain to what extent the safety and low cost that are advantages of aqueous batteries can be preserved with this new electrolyte design.
There is still a distance between what computation shows to be possible and what experiments prove to be real. The work of measuring that distance is only just beginning.