Moving atoms could make globally controlled quantum computers programmable

Scaling a quantum computer requires more than adding qubits. It also requires performing the right operations on selected qubits while leaving others undisturbed. Locally focused light can provide that selectivity, but maintaining precise optical control across changing gate patterns creates a calibration burden.

In an arXiv preprint, Zhihao Zhang, Shuaifan Cao and colleagues propose a different arrangement: use global pulses for quantum operations and move auxiliary atoms to determine which qubits interact. Their work presents an architecture and numerical benchmarks, not an experimental demonstration of a complete fault-tolerant processor.

Selecting qubits by moving their neighbours

The design combines stationary caesium atoms, which store quantum information, with mobile rubidium atoms. Optical tweezers move the rubidium atoms, allowing them to serve either as operation selectors or as auxiliary qubits for detecting errors.

The mechanism relies on Rydberg states—highly excited atomic states that interact strongly. Bringing a rubidium atom sufficiently close to a chosen caesium atom enables a conditional operation through the Rydberg blockade interaction. Global fields illuminate the relevant species, but the arrangement of nearby atoms determines the intended interacting pairs.

The authors package these capabilities into the Global Rydberg Atom Instruction Language, or GRAIL. Its instructions cover global rotations, rubidium transport, Rydberg pulses, state preparation and measurement. By combining them, they construct a universal physical gate set: H, T and CZ. Together, these operations can approximate arbitrary quantum computations.

For a two-caesium-qubit gate, a shared rubidium atom interacts sequentially with both qubits, mediating their operation before returning to its initial state. Ideally, the compiled sequences leave unselected atoms unchanged.

Testing errors under global illumination

Real global pulses also affect idle atoms. This unwanted disturbance, called crosstalk, is a central challenge for the proposal.

The researchers simulated Rydberg dynamics including spontaneous decay and finite interaction strength. They then converted the resulting errors into a circuit-level noise model for testing error-correcting protocols. At a Rydberg driving frequency of 2.38 megahertz, the predicted fidelities of the compiled H, T and CZ gates all exceed 99.5%. Stronger driving reduces the calculated crosstalk; at 12 megahertz, crosstalk infidelities for these sequences are below 0.1%.

These figures describe modelled gate performance, not measurements from a working implementation.

Protecting stored quantum information

Next, the team translated a ZXXZ surface code into GRAIL. A surface code spreads one logical qubit across many physical qubits and repeatedly measures checks that reveal evidence of errors. Here, moving rubidium auxiliary atoms collect those checks from the caesium data register.

The proposed schedule uses eight parallel transport layers per checking round, independent of code distance—the parameter describing the code’s protection strength.

Simulations including crosstalk give a fitted memory threshold of 1.43%. Below this threshold, increasing code distance reduces logical errors. Importantly, the percentage refers to the model’s chosen physical-error parameter: the leakage probability of a native rubidium–caesium CZ gate, not a uniform error probability for every operation.

At 2.38 megahertz, increasing distance by two suppresses the logical error per cycle by about 8.76 times. Extrapolating the simulated trend, the authors estimate that distance 17 could bring memory errors below one in a billion per cycle. That larger-distance result is a projection rather than a directly sampled benchmark.

Beyond memory—and within the model

Universal fault-tolerant computation also needs encoded operations and special resource states. The authors compile logical Clifford operations and adapt magic state cultivation, which prepares resource states for logical T gates. Parallel execution reduces global-pulse exposure, while postponed measurements accommodate global rubidium readout.

Using a logical-Y cultivation benchmark to estimate the target magic-state protocol, they report a logical error rate of 1.98 × 10⁻¹⁰ at 11.9 megahertz, requiring an average of 2.43 attempts per accepted output. This result includes rejecting attempts that fail specified checks or reliability criteria.

The predictions depend on modelling approximations: leakage and crosstalk are represented by depolarizing channels in the circuit simulations, and preparation, measurement and microwave errors receive assigned probabilities. The proposal therefore establishes a model-based route to simpler gate addressing, not proof that all control challenges have disappeared. Its central trade is clear: replace locally calibrated gate beams with carefully engineered atom positions.