Source-linked AI summary
Aquila: QuEra's 256-qubit neutral-atom quantum computer
Jonathan Wurtz, Alexei Bylinskii, Boris Braverman, Jesse Amato-Grill, Sergio H. Cantu, Florian Huber, Alexander Lukin, Fangli Liu, Phillip Weinberg, John Long, Sheng-Tao Wang, Nathan Gemelke, Alexander Keesling
TL;DR
Aquila addresses the need for practical evidence about the capabilities and limits of near-term neutral-atom quantum hardware. The whitepaper presents Aquila as a programmable analog FPQA, explains its operation, and demonstrates applications from single-qubit dynamics to combinatorial optimization. Its examples include entangling dynamics and optimization workflows, while the authors stress that analog-mode performance should not be interpreted as future gate fidelity.
Problem
Near-term quantum devices have limited generality and problem size, motivating a transparent account of Aquila’s strengths and limitations.
Method
The whitepaper explains Aquila’s analog FPQA architecture and evaluates it through benchmarks and five progressively complex examples implemented on the device.
Results
The demonstrations include entangling dynamics with single-cluster T2 times of 8.9, 6.9, and 6.6μs for N=2, 3, and 4, respectively, plus a neutral-atom maximum-independent-set workflow.
Takeaways & Limitations
Aquila supports programmable analog quantum dynamics for tasks such as simulation and optimization, with performance interpreted within its noisy, analog operating mode.
Takeaways & Limitations
Return probabilities from the Levine–Pichler analogue are not gate fidelities because Aquila operates in analog mode without the hyperfine-state gate implementation.
Abstract
from arXiv · showhide
The neutral-atom quantum computer "Aquila" is QuEra's latest device available through the Braket cloud service on Amazon Web Services (AWS). Aquila is a "field-programmable qubit array" (FPQA) operated as an analog Hamiltonian simulator on a user-configurable architecture, executing programmable coherent quantum dynamics on up to 256 neutral-atom qubits. This whitepaper serves as an overview of Aquila and its capabilities: how it works under the hood, key performance benchmarks, and examples that demonstrate some quintessential applications. This includes an overview of neutral-atom quantum computing, as well as five examples of increasing complexity from single-qubit dynamics to combinatorial optimization, implemented on Aquila. This whitepaper is intended for readers who are interested in learning more about neutral-atom quantum computing, as a guide for those who are ready to start using Aquila, and as a reference point for its performance as an analog quantum computer.
Selected “best practices.”
Aquila best practices emphasize smooth, fast, noise-robust protocols, appropriate shot counts, valid atom arrays, and geometry choices that avoid sensitivity to hardware limits.
- Implementation considerations: Post-select correctly filled arrays, parallelize repeated few-atom configurations, disable the Rabi drive during large phase jumps, and place atoms deep within or far from the blockade radius.These practices address array validity, shot reduction, AOM behavior, and sensitivity to thermal position fluctuations.
- On designing smooth waveforms: Smooth waveforms help avoid unexpected behavior caused by the finite bandwidth of Aquila’s optical control elements.Rapidly varying waveforms can exceed the effective control bandwidth.
- On maximizing Rabi frequency: Use the maximum possible Rabi drive when fixed pulse area makes shorter protocols desirable for reducing decoherence.For fixed Ωt, increasing Ω reduces the protocol duration.
- On the robustness of adiabatic protocols: Adiabatic protocols are relatively insensitive to phase, amplitude, and position noise, making them a robust analog-computation option.The recommendation applies when designing analog algorithms.
- On choosing the number of shots: Choose shot counts according to the trade-off between noise and speed or cost, with about 100 measurements as a practical middle ground.High-resolution phase diagrams and low-probability outcomes may require up to 1000 shots, while parallelized few-atom arrays may use as few as 25.
1. Introduction
The whitepaper presents Aquila as a 256-qubit neutral-atom FPQA and analog Hamiltonian simulator, documenting its operation, benchmarks, and applications while emphasizing both quantum effects and current limitations.
- Device and scope: Aquila is a field-programmable qubit array and analog Hamiltonian simulator that executes programmable coherent dynamics on up to 256 neutral-atom qubits.It is available through Amazon Braket on AWS.
- Whitepaper scope: The whitepaper combines an overview of Aquila’s operation with key performance benchmarks and five examples ranging from single-qubit dynamics to combinatorial optimization.Associated Jupyter notebooks implement the examples through the Braket SDK.
- Intended audience: The document targets readers learning neutral-atom quantum computing, users preparing to run Aquila, and readers seeking a performance reference for analog quantum computing.It assumes basic quantum-mechanics knowledge but no further advanced concepts.
- Positioning and limitations: Aquila is presented transparently as a noisy, decohering device with limitations, while the whitepaper reports evidence of quantum effects underlying its operation and performance.The stated goal is to highlight both strengths and limitations through data.
- Data basis: The reported dataset comprises 96,250 measurements across 1,884 tasks, acquired on Aquila on June 9, 2023.The measurements support the examples presented in the whitepaper.
1.1. Background and literature
The background places Aquila within rapid neutral-atom progress, highlighting demonstrations from Rydberg-qubit experiments to entanglement, quantum phases, and optimization, while illustrating its compact optical hardware.
- Field development: Neutral-atom quantum computing has advanced from early platform proposals and Rydberg-qubit proof-of-concept experiments to increasingly capable demonstrations.The cited advances include high-fidelity gates, large entangled states, quantum phases, topological phases, and optimization.
- Aquila hardware: Aquila’s interior uses optical elements, lasers, and cameras focused on a vacuum cell containing dilute rubidium gas to control up to 256 qubits in a very small region.The controlled area is described as less than three human hairs wide.
- Industry context: The field’s progress has supported the creation of multiple neutral-atom quantum-computing companies, including QuEra and several other firms.The passage attributes these companies to the success of academic research laboratories.
- Further reading: For broader technical background, the whitepaper recommends review articles by Saffman and Morgado.These references are offered as supplementary neutral-atom quantum-computing resources.
1.2. The key ingredients of neutral-atom quantum computing
Aquila combines Rb-87 atoms, reconfigurable optical-tweezer arrays, laser control, and Rydberg interactions into a programmable analog quantum computer. Its architecture supports arbitrary geometries and coherent dynamics, while measurement, loading, and spacing constraints define practical operating boundaries.
- Key ingredient: Rubidium atoms: Aquila uses individual Rb-87 atoms as qubits, with electronic states manipulated by lasers and read out through state-dependent fluorescence.The platform uses ground-Rydberg and hyperfine qubits; Aquila currently operates with the ground-Rydberg qubit in analog mode.
- Key ingredient: Rydberg states and the Rydberg blockade: Rydberg blockade prevents a nearby second atom from reaching the Rydberg state, enabling robust interaction-based entanglement.Outside the blockade radius, the second atom can be driven; inside it, strong interactions detune the transition.
- Key ingredient: Rubidium atoms: Measurements are destructive because atoms detected in the Rydberg state are lost, so each experimental cycle rebuilds the array.The process is relatively slow, below 10 Hz, but permits a different geometry to be selected from shot to shot.
- Key ingredient: FPQA with arbitrary geometry and optical tweezers: The array supports structures such as regular registers, Kagome lattices, and coastline-shaped arrangements for quantum simulation and geographical optimization.Positioning remains constrained by a 75 μm × 76 μm area, minimum 4 μm site spacing, and row-based sorting requirements.
- Key ingredient: FPQA with arbitrary geometry and optical tweezers: Aquila’s field-programmable architecture uses optical tweezers to position and rearrange up to 256 atoms in user-configurable geometries.An SLM creates quasi-static trap layouts, while AODs move atoms on microsecond timescales for sorting and rearrangement.
- Key ingredient: Photonics, lasers, and analog control: Aquila’s analog programs specify time-dependent Rabi drive, detuning, phase, and atom positions rather than gate sequences.Always-on interactions can build entanglement and correlations quickly, but analog operation is specialized rather than necessarily universal.
1.3. The Rydberg Hamiltonian
Aquila’s analog dynamics combine laser-driven single-atom transitions with state-dependent van der Waals interactions between atoms. Programs specify time-dependent drive amplitude, phase, detuning, and atom positions, with evolution generated by a time-dependent Hamiltonian.
- Aquila’s dynamics combine laser-driven ground–Rydberg transitions with state-dependent van der Waals interactions between neighboring atoms.The resulting evolution is generated by a time-dependent Hamiltonian.
- Four controls define a quantum program: Rabi amplitude Ω(t), drive phase ϕ(t), detuning Δ(t), and atom positions x⃗_i.These controls determine single-qubit rotations, resonance offset, and Rydberg–Rydberg interaction strengths.
- The ground and Rydberg states encode logical 0 and 1, while n̂_i counts Rydberg excitations and measurements use only the logical Z basis.
- Aquila’s conventions use radians per microsecond and micrometers, whereas Amazon Braket uses radians per second and meters.The Braket convention differs by a factor of 10^6 in the relevant units.
1.4. Dominant sources of error
Aquila’s fidelity decreases with evolution time because coherent and incoherent noise accumulate. Dominant errors arise from laser fluctuations, atom motion, state decay and scattering, spatial inhomogeneity, and imperfect measurement.
- Longer evolution times reduce state fidelity as multiple coherent and incoherent noise sources accumulate.
- Short protocols and maximum available Ω help limit decoherence for a fixed pulse area Ωt.
- Laser phase and amplitude noise produce shot-to-shot and time-dependent variations in Ω and Δ, causing averaged expectations to approach their time averages.
- Thermal atom motion creates Doppler-induced detuning variance and is most sensitive when detuning and Rabi frequency are similar.
- State decoherence and scattering cause incoherent decay through losses involving the intermediate and Rydberg states.
- Spatial inhomogeneity changes Ω and Δ across the array, while imperfect retrapping can misclassify ground and Rydberg states during measurement.
1.5. Datasheet of Aquila capabilities and performance metrics
Aquila’s datasheet reports calibrated limits for geometry, preparation, measurement, Hamiltonian control, and coherence. Representative benchmarks include 7.5 μsec driven-qubit decoherence, 8.9 μsec for a blockaded pair, and correlation lengths of 3.6 and 5.7 sites in one- and two-dimensional states.
- The listed program restrictions are chosen within Aquila’s expected performance limits, although some demonstrated experiments exceed them.Examples include coherent evolution to 10 μsec and geometries with 115 μm vertical height.
- Position, drive, detuning, filling, and detection specifications quantify Aquila’s state-preparation, Hamiltonian-control, and measurement errors.Examples include 0.050 μm systematic position error, 0.007 filling failure probability, and 0.08 Rydberg mis-detection probability.
- Qubit dephasing times are 5.8 μsec without drive and 11.4 μsec for incoherent processes measured with spin echo.
- 7.5 μsec is the driven-qubit decoherence time T2^Rabi, compared with 8.9 μsec for an isolated pair of mutually blockaded qubits.Both values include coherent and incoherent processes.
- 3.6 sites is the correlation length of an adiabatically prepared Z2 state in one dimension.
- 5.7 sites is the reported correlation length of an adiabatically prepared checkerboard state.
2. Example 1: Single-qubit dynamics
Aquila demonstrates single-qubit analog control through programmable amplitude, phase, detuning, and waveform schedules. Rabi, Ramsey, Floquet-inspired, and spin-echo protocols show coherent control while exposing finite-bandwidth, inhomogeneity, noise, and decoherence limits.
- 2. Example 1: Single-qubit dynamics: Ω(t) and Δ(t) can prepare any single-qubit superposition because their integrated values determine the Bloch-sphere angles θ and ϕ.
- 2.1. Rabi oscillations: Resonant Rabi driving oscillates between |0⟩ and |1⟩, whereas detuning produces oscillations between |0⟩ and a superposition state.
- 2.1. Rabi oscillations: Evolution shorter than 100 ns is accessible, but longer-time oscillations decay because of coherent and incoherent noise and decoherence.
- 2.1. Rabi oscillations: 7.5 μsec is the average single-atom Rabi coherence time, while ensemble averaging gives approximately 3.6 μsec under on-resonant oscillations.The ensemble value includes dephasing from slightly inhomogeneous Rabi drive.
- 2.2. Time-dependent protocols: A Ramsey protocol varies the hold time between two rotations to calibrate resonance and measure phase coherence, yielding T2* values near 5.5–5.8 μsec.
- 2.2. Time-dependent protocols: A Floquet-inspired protocol holds Ω = 15 rad/μs while varying detuning as Δ = 15sin(15t), demonstrating nontrivial coherent evolution from arbitrary waveforms.
- 2.3. Dynamical decoupling protocols: Spin echo uses phase-shifted rotations to reverse accumulated phase and cancel environmental phase noise to first order.
3. Example 2: Many-qubit dynamics
Aquila’s many-qubit dynamics use Rydberg interactions and blockade to control collective excitations, prepare entangled W states, and probe blockade-radius transitions. Experiments demonstrate √N Rabi enhancement, while finite-time errors, decoherence, thermal motion, and nonblockaded dynamics limit agreement with ideal behavior.
- Rydberg blockade: The Rydberg blockade excludes doubly excited states within a distance-dependent radius, while interactions rapidly weaken as R^-6.At 4 μm, the interaction is about 1,320 rad/μsec, whereas at 16 μm it is about 0.32 rad/μsec.
- Adiabatic state preparation: Adiabatic preparation switches from one to two Rydberg excitations across the static blockade radius, with finite-time errors and noise broadening the transition.The protocol ramps Ω and Δ between initial and target Hamiltonians; the ideal transition is sharp only in the adiabatic limit.
- Rabi frequency enhancement: √N enhancement increases the effective Rabi frequency for N blockaded atoms by coupling the ground state to a symmetric single-excitation W state.Only the ground state and symmetric single-excitation state participate in the relevant low-energy dynamics.
- Rabi frequency enhancement: Larger atom numbers reduce separation between energy scales, producing nonblockaded dynamics that deviate from the ideal collective model.This limitation is visible in the many-atom Rabi-oscillation results.
- Levine-Pichler gate analogues: The Levine-Pichler analogue reaches 97.9% and 96.0% ground-state return probabilities for one and two atoms, but these are only weak upper bounds on gate fidelity.Aquila’s analog-mode operation and the fixed, unoptimized protocol mean the return probabilities are not direct indicators of future hardware gate fidelities.
4. Example 3: Many-body ordered phases
Aquila prepares ordered many-body phases through adiabatic evolution, producing 1D Z2 states and distinct 2D checkerboard and striated phases. Measurements show finite but substantial correlations, including correlation lengths of ≈3.6 sites in 1D and ≈5.7 sites in 2D.
- 4.1. The 1D Z2 phase: ≈58% of measurements yielded the target 1D Z2 bitstring for 11 atoms after 4 μs, with alternating Rydberg occupation across sites.The observed probability is consistent with an ≈8% error in detecting Rydberg states.
- 4.1. The 1D Z2 phase: The prepared 1D Z2 state has a correlation length of ≈3.6 sites and clear antiferromagnetic order.The connected correlation function is fit by an exponential decay with distance.
- 4.2. Adiabatic preparation performance characterization: At fixed 7 μs evolution time, the Z2-state probability falls exponentially with chain length but remains ≈40% for N=19.The decline is attributed to finite correlation length and measurement error.
- 4.2. Adiabatic preparation performance characterization: Fidelity does not decay significantly for evolution times of order 10 μs, indicating coherent evolution within the 4 μs window accessible through Amazon Braket.Noiseless theory predicts increasing preparation probability with time, while decoherence eventually limits longer evolution.
- 4.3. The 2D striated and checkerboard phase: Changing the 2D lattice spacing produces checkerboard and striated phases, with intermediate Rydberg densities in the striated phase indicating quantum fluctuations.The two phases were prepared on an 11 × 11 grid using adiabatic evolution with different lattice spacings.
- 4.3. The 2D striated and checkerboard phase: The 2D checkerboard state exhibits long-range correlations with a correlation length of ≈5.7 sites despite noise and decoherence.The connected correlation function decays exponentially with separation.
5. Example 4: Many-body quantum scars
Aquila simulates many-body scar dynamics by adiabatically preparing a Z2 state and then quenching into non-equilibrium evolution. The resulting Rydberg-density oscillations and Néel-state revivals persist instead of rapidly thermalizing, consistent with scar behavior.
- 5.1. Protocol: A two-part protocol first adiabatically prepares a Z2 state and then quenches the system under scar dynamics.The quench is implemented by setting the detuning to zero.
- 5.2. Scar dynamics: Persistent Rydberg-density oscillations after the quench indicate coherent evolution rather than rapid thermalization.Aquila measurements agree well with classical simulation results for the density dynamics.
- 5.2. Scar dynamics: The tracked Néel-state probability shows clear revivals at fixed times after the quench, indicating repeated large overlap with the initial ordered state.These revivals are consistent with the quantum many-body scar picture.
6. Example 5: Maximum independent set on unit disk graphs
Aquila encodes maximum independent set on unit disk graphs by matching the graph’s unit-disk radius to the Rydberg blockade radius and preparing the system adiabatically. On one graph, the hybrid method occasionally reaches the MIS, but across 50 graphs it shows no ensemble advantage over the classical baseline.
- Problem encoding: The MIS problem seeks the largest vertex subset with no connected pair, and is encoded by placing atoms at graph positions with R_ud=R_b.The ground state is measured through the atoms excited to the Rydberg state.
- Post-processing: For small final detuning, the Rydberg count is below the MIS because the effective blockade radius is large; at large detuning, independent-set violations increase.Post-processing removes violations and greedily adds vertices to produce maximal independent sets.
- Protocol: The average maximal-independent-set size is optimized near Δf≈40.The hybrid and classical distributions are compared after post-processing.
- Single-graph performance: The hybrid algorithm averages 57.5 mIS vertices versus 58.0 for the classical algorithm, but has about a 4.5% chance of finding the MIS of size 60.The comparison uses 200 machine shots and 10 rounds of post-processing per shot.
- Ensemble performance: ⟨PR⟩=0.974 across 50 graphs, indicating no ensemble advantage for the hybrid algorithm over the classical-only baseline.Only a few atypical instances showed any performance boost.
About QuEra
QuEra is a Boston-based neutral-atom quantum-computing company founded in 2018. It develops scalable quantum computers aimed at useful problems that are classically intractable, including through its Aquila machine.
- About QuEra: QuEra Computing is located in Boston and makes advanced quantum computers based on neutral atoms.
- About QuEra: The company was founded in 2018 and builds on research conducted at Harvard University and MIT.
- About QuEra: QuEra aims to build scalable quantum computers for useful, commercially relevant problems that are classically intractable.