Google Uses AI to Design Less Demanding Quantum Programs

Google researchers report that artificial intelligence (AI) teams have designed less demanding quantum programs for scientific calculations involving fusion and chemistry. Their preprint, first posted November 6 and revised November 9, describes a way to improve programs while the machines needed for these calculations are still being developed. [1]

The reported resource reductions are approximately 11 times for a fusion calculation, eight times for photochemistry and 2.5 times for drug-metabolism research. The comparisons measure the demands of proposed programs against earlier designs. For scientists, lowering those demands could bring a useful calculation within reach of a smaller future machine, provided the calculation achieves the accuracy its scientific purpose requires. [1]

Human researchers specified the tasks. AI teams searched for programs and supplied mathematical proofs that Lean, a proof-checking system, checked against those requirements. A separate checker guarded the original problem statement and permitted assumptions. Together, those checks address a practical concern in automated research: a proposed improvement has to perform the calculation that scientists asked for. [1, 6]

The applications remain ambitious. The authors write, “Our end-to-end simulations still demand far more resources than projected for the first fault-tolerant quantum computers.” Such computers would use error correction to protect calculations from faulty operations. The result establishes improved program designs and author-reported resource estimates; making them useful in a laboratory also requires suitable hardware and comparisons with the best conventional methods. [1]

A quantum computer stores information in physical qubits, which are susceptible to errors. Error correction combines many of them into more reliable logical qubits. Google’s metric multiplies logical qubits by T gates, a costly type of computing operation, to capture two demands on a machine. Converting that score into hours or dollars requires a hardware design, operating speeds and an error-correction plan. A 2022 study by Michael Beverland and colleagues estimated that selected applications could require hundreds of thousands to millions of physical qubits under particular assumptions. That framework shows how software savings and device improvements jointly shape the scale of a useful computer. [1, 2]

The fusion calculation concerns stopping power: how quickly charged particles transfer energy to fuel. Better estimates could help scientists interpret experiments and design fusion targets. Nicholas Rubin and colleagues published an earlier quantum approach in 2024. Photochemistry asks how molecules change when light sets their electrons and nuclei in motion. Some reactions require following those motions together, a difficult calculation that Matthew Pocrnic and colleagues addressed in a February 2026 paper. These scientific models supply concrete problems against which researchers can judge proposed computational savings, their accuracy and their value. [3, 4]

Lean checks a formal chain of reasoning against stated definitions and assumptions. Comparator tooling adds checks that a submitted proof addresses the original statement and uses only permitted assumptions, within a trustworthy checking environment. Scientists still choose the physical model and judge whether it captures the phenomenon they want to study; hardware engineers establish device reliability. The November 9 manuscript says the Lean library, problem specifications and verified solutions will be released shortly. Making those materials available would let other researchers inspect and build on the reported designs. [1, 5, 6]

An October report from Anthropic offers a related example of AI searching for better quantum circuits with an automated evaluator. Across eight tasks it reported improved designs; in one comparison with the smallest previous design, a circuit used 17 times fewer gates and 12 percent more qubits. That trade-off shows why designers track both operations and memory. The two projects give scientists and engineers a larger set of candidates to examine. Their practical value will depend on translating verified program improvements into dependable scientific calculations at an affordable hardware scale. [7]

Sources

[1] Shirobokov and colleagues, Discovering formally verified quantum programs with AI. Version 2, November 9, 2026.

[2] Beverland and colleagues, Assessing requirements to scale to practical quantum advantage. November 2022.

[3] Rubin and colleagues, Quantum computation of stopping power for inertial fusion target design. PNAS, 2024.

[4] Pocrnic and colleagues, Efficient Simulation of Pre-Born-Oppenheimer Dynamics on a Quantum Computer. February 2026.

[5] Lean, official introduction to the programming language and proof assistant.

[6] Lean Comparator, official documentation and checking assumptions.

[7] Anthropic, Designing more efficient quantum circuits with Claude. October 16, 2026.