AQVolt26: how physics-based AI is rewriting battery discovery

Large Quantitative Models are quietly becoming the foundry where next-generation battery chemistry gets forged. SandboxAQ's AQVolt26 release is the clearest signal yet that materials discovery has moved from the lab bench to the compute cluster. The U.S. Commerce Department then put $500 million behind the same idea.

The pitch is simple to state and hard to deliver. Battery performance is a materials problem. Finding the right solid electrolyte means searching a space with more plausible chemistries than any team could synthesize by hand. AI changes the economics of that search.

Two fields that were never supposed to share a pipeline

Materials science and artificial intelligence grew up in different buildings. One counted atoms. The other counted tokens. For most of the last decade the connection was ceremonial: a neural network would predict a property, a chemist would shrug, and the result would sit in a preprint nobody synthesized. The divide was not ignorance. It was the cost of moving from a prediction to a physical sample, a cost neither field could lower on its own until compute got cheap enough to simulate the chemistry directly.

That gap is closing because the bottleneck moved. The expensive part of modern battery research is no longer guessing a candidate. It is simulating how a candidate behaves when lithium ions try to move through it at the temperatures a real cell reaches. Classical Density Functional Theory (DFT) can do this, but a single dynamic simulation of a soft halide lattice can cost more compute than a small training run. Scale that to the millions of structures worth screening and the math breaks.

Large Quantitative Models (LQMs) are the bridge SandboxAQ built. Where a Large Language Model trains on text, an LQM trains on rigorous physical and chemical data: quantum calculations, force fields, spectra. The model learns the shape of energy itself, not the shape of sentences. The company frames this as a horizontal platform, but the battery work is where the claims get specific.

What AQVolt26 changes about the search

AQVolt26 is a dataset and a suite of machine-learning interatomic potentials (MLIPs) released by the AISim team in April 2026. The headline number is 322,656 high-fidelity DFT calculations of lithium halide electrolytes, all run at the r²SCAN level of theory. r²SCAN is a meta-GGA functional that captures the awkward mid-range dispersive interactions inside soft halide lattices far better than cheaper approximations. The dataset maps the high-temperature, distorted atomic configurations that general-purpose models tend to get wrong.

The problem it attacks is concrete. Halide solid electrolytes are promising: high ionic mobility, wide stability, and the mechanical give needed to hold a solid interface together. But their anions are so polarizable that at the temperatures above 1,000 K required to simulate ion transport, the atoms distort wildly. Foundational potentials trained near equilibrium quietly fail in that regime. AQVolt26 was built to close that blind spot, and when co-trained with existing open datasets such as MatPES and MP-ALOE it keeps physical consistency where the older models drift.

The released eSEN models were stress-tested under extreme lattice deformations of ±20 percent on the public MLIP Arena benchmark. They posted a 0.2 percent failure rate with perfect monotonic energy scaling, well ahead of the baselines it was measured against. The models and dataset are published on Hugging Face, so the work is reproducible rather than demoed.

This is not the first time AI has pointed at a battery material. As we wrote in August, Cornell's IonNet surfaced 87 new solid-state electrolyte candidates by reasoning from chemistry alone, without synthesizing a single cell. AQVolt26 is the same instinct aimed at a harder layer: the dynamics of those materials at the temperatures a real battery actually runs.

Why solid-state, and why the grid is watching

All-solid-state batteries (ASSBs) promise two things liquid cells cannot: energy density high enough to change vehicle range math, and the removal of the flammable organic solvent that makes lithium-ion dangerous under abuse. For the grid, the second property matters as much as the first. Long-duration storage that does not carry a thermal-runaway liability is easier to site, insure, and stack next to homes and data centers.

The catch is the interface. A solid electrolyte has to stay in contact with a solid electrode through thousands of charge cycles while ions shuffle across the boundary. That is a materials problem with no shortcut, and it is exactly the regime where simulation has to be right at temperature, not at zero Kelvin. AQVolt26's value is letting researchers screen ionic conductivity across novel chemistries without burning months of compute per candidate.

The field context matters. Microsoft and the Pacific Northwest National Laboratory used AI to narrow 32 million candidate materials to a short list in roughly 80 hours, then flagged a lithium-reducing solid-state electrolyte worth building. That result drew attention because it compressed a search that used to take years into a work week. AQVolt26 is the next step: not just picking from a list, but modeling the behavior that decides whether a pick survives contact with an electrode.

The half-billion-dollar vote

In June 2026 the U.S. Commerce Department signed a definitive agreement awarding SandboxAQ $500 million through the CHIPS R&D program. The stated target is not a better chip. It is the material stack underneath the chip: PFAS-free process chemicals, fabrication catalysts, rare-earth-free magnets, and battery systems for semiconductor facility backup power. In connection with the award, Commerce takes a minority non-voting equity stake.

That framing is the part investors should read twice. Washington is treating AI-enabled materials discovery as infrastructure, on the same ledger as fabs. The bet is that a repeatable simulation-to-lab pipeline can harden supply chains that foreign suppliers have dominated for decades. its public test is specific: show that physics-based AI produces candidates which are not merely computationally interesting but manufacturable, with American partners, across several material classes.

The materials informatics sector around this work is still small, but the number of startups selling the search rather than the cell is climbing. Most will not build batteries. They sell the pipeline that finds the chemistry, and that is the layer it is trying to own.

Where the model meets the lab

The honest limit is the one every materials-AI paper states and every press release forgets. A model that ranks ionic conductivity does not prove a cell will cycle. The distance from a favorable potential-energy surface to a coin cell that survives 1,000 charges is still measured in years of synthesis, characterization, and failure.

Domain data is the deeper constraint. AQVolt26 is a targeted complement, not a replacement, for foundational datasets. Its strength is the high-temperature regime general models miss, which means its usefulness depends on being co-trained with near-equilibrium data rather than standing alone. A model tuned to one family of halides does not automatically generalize to sulfides or oxides, and the paper is explicit that near-equilibrium relaxation data helps in task-specific ways rather than universally.

Intellectual property is the quiet gate. The companies sitting on the best datasets treat them as crown jewels. the decision to open AQVolt26 is unusual and strategically interesting: it builds a commons around its models while the commercial moat shifts to the LQM platform and the CHIPS-funded application pipeline. Rivals who keep their data closed may win specific deals and lose the standards war.

The economics of validation are the part models cannot skip. Every proposed candidate still has to be made and tested by people, and a single failed synthesis can cost more than a thousand simulations combined. The pipeline's real value is not candidates per second but successful validations per dollar of screening, because the bottleneck has moved from asking the question to trusting the answer. A model that surfaces ten strong halides is useful only if the lab can afford to build and cycle even two of them. That ratio, not the benchmark failure rate, is what will decide whether AQVolt26 pays for itself inside a battery program, and it is the number no preprint can show.

The case against the hype

Skeptics have a fair point that the field keeps announcing the end of trial-and-error while trial-and-error quietly continues. Screening millions of candidates fast is real. Validating even a handful in a working cell is still slow, still expensive, and still where most ideas die. A 0.2 percent failure rate on a static benchmark is not the same as a 0.2 percent failure rate in a pilot line.

There is also a concentration risk. If every materials-AI effort leans on the same foundational datasets and the same meta-GGA functionals, the field can converge on a narrow slice of chemistry that looks safe to models and forgets the corners where breakthroughs live. AQVolt26's answer is more targeted sampling, not less. Whether that is enough to keep discovery honest is an open question the next few validated cells will answer.

The counter-case is not that AI is overstated. It is that the timeline is. Discovery is accelerating. Translation is not. The investor who prices AQVolt26 as a battery in 2027 will be disappointed. The investor who prices it as a permanent reduction in the cost of asking nature a hard question has the right horizon.

How the models actually run

A machine-learning interatomic potential is a learned stand-in for a quantum calculation. Train it on thousands of labeled structures where the true energy and forces are known, and it predicts them for new structures in microseconds instead of hours. the eSEN models extend this to the anharmonic regime where halide atoms wander far from their rest positions, the exact condition that breaks cheaper potentials and the exact condition a hot battery creates.

The payoff is throughput. Once the potential is fast, molecular dynamics can simulate ion motion across millions of candidate lattices and rank them by ionic conductivity without a single DFT step per structure. AQVolt26's contribution is the training data for that high-temperature regime: surrogate-driven exploration of distorted configurations so the model does not silently fail when a real cell heats up. The arXiv paper is explicit that near-equilibrium data helps, but as a task-specific complement rather than a universal cure, and that is an honest limit most marketing skips.

The field is no longer a solo act

It is not alone in betting on simulation-first discovery. Microsoft and the Pacific Northwest National Laboratory showed the headline version of the idea, compressing a 32-million-candidate search to a short list in about 80 hours and surfacing a lithium-reducing solid-state electrolyte worth building. The difference now is that the tooling is opening. AQVolt26 ships as models and data anyone can co-train, which turns a vendor advantage into a commons and pressures closed rivals to show their own validation.

For investors the pattern matters less in the lab and more in the stack. Whoever owns the loop from simulation to validated cell, and the data that trains it, sits closest to the margin. The $500 million federal award is less a bet on one battery than a bet that this loop is infrastructure, and that the country that runs it controls the materials underneath the chip.

What a validated win looks like

The claim becomes real only when a cell cycles. The first milestone to watch is an external lab building a coin cell from an AQVolt26-proposed halide and reporting ionic conductivity and cycle life that hold up outside SandboxAQ. A second milestone is a battery or automotive OEM citing the models in a materials decision, not just a paper. A third is the CHIPS program's own scorecard: which of its four material tracks reaches a domestic manufacturing partner first, and on what timeline.

None of these is years away in principle, but all of them are where materials AI has stalled before. The pattern of the last decade is a dazzling simulation followed by a quiet pilot that never scales. AQVolt26 changes the search speed. It does not change the fact that a solid electrolyte still has to survive contact with a real electrode for thousands of cycles. The investor who watches cells, not benchmarks, will see the signal first, and the signal will arrive in a lab notebook rather than a press release.

What to watch

The story is no longer whether machines can propose battery materials. They can, at scale, with numbers now reproducible by outsiders. The variable that decides value is the lab's ability to convert proposals into cells, and the policy's ability to convert awards into domestic supply. It has the model, the open data, and the federal check. What it does not yet have is the proof that the pipeline end to end beats the bench.

For a reader evaluating private positions in deep tech, the cleaner read is structural. Materials discovery is becoming a software business with a physical tail. The margins will sit with whoever owns the loop, not the lookup.