Quantifying the Dual-isotope Advantage for Ytterbium-array Surface Codes using Realistic Noise Models

AI summary

Overview

Research area: Quantum error correction on neutral-atom quantum computers — specifically ytterbium (Yb) tweezer arrays and surface-code quantum memories.

Technical level: Advanced. The paper assumes familiarity with stabiliser codes, syndrome extraction, Pauli twirling, and Rydberg-blockade gates.

Scope (one sentence): The paper builds an experimentally motivated, Clifford-simulatable noise model for dual-isotope $^{171}$Yb–$^{174}$Yb arrays and uses it to compare the surface-code logical error rate of dual-isotope in-place measurement against single-isotope shelving, zoned measurement, and SWAP-based syndrome extraction.

What This Paper Is About

Neutral-atom processors are slower than superconducting ones largely because ancilla atoms must be physically moved, shelved, or otherwise manipulated to be measured, which adds time overhead to every error-correction cycle. A dual-isotope array sidesteps this: because the two isotopes respond to different frequencies, ancilla atoms can be read out where they sit while data atoms stay dark, giving constant-time syndrome extraction independent of code distance. The question the paper asks is whether this architectural convenience actually translates into better logical performance once realistic noise is included, and which physical error sources dominate the result.

Key Contributions

1. A realistic noise model for dual-isotope Yb. The authors construct a Clifford-compatible noise model for $^{171}$Yb–$^{174}$Yb systems using generalised Pauli twirling, based on effective four-level ($^{174}$Yb: ${|0\rangle,|1\rangle,|r\rangle,|L\rangle}$) and six-level ($^{171}$Yb: ${|m_0\rangle,|m_1\rangle,|g_0\rangle,|g_1\rangle,|r\rangle,|L\rangle}$) descriptions with explicit effective loss states.

2. An open-source simulator. The model is released as DualYbSim, a Python library that wraps stim and inserts the twirled noise channels into Clifford circuits.

3. A head-to-head architectural comparison. Rotated and XZZX surface-code memories are simulated across dual-isotope in-place measurement, single-isotope shelving, single-isotope zoned measurement, and both normal and SWAP syndrome extraction.

4. An error-budget decomposition. Using weighted quadratic fitting (solved with CVXPY), the authors split the inverse error-suppression factor $\Lambda^{-1}$ into per-error-source contributions $C_i$, identifying which physical imperfections limit the code.

Main Findings

• Dual-isotope wins on logical error rate. The ($^{171}$Yb-g, $^{174}$Yb) configuration with normal syndrome extraction and in-place measurement achieves the lowest logical error rate per round among all configurations considered, for both rotated and XZZX codes, and shows the steepest improvement with code distance.

• Quantified suppression advantage. The dual-isotope architecture has $\Lambda^{-1} \approx 0.54$, whereas the other configurations range approximately from 0.62 to 0.72 — a smaller $\Lambda^{-1}$ meaning stronger error suppression per unit distance.

• Rydberg decay dominates the error budget. Decay from the Rydberg state during two-qubit gates contributes 80% for the ($^{171}$Yb-g, $^{174}$Yb) dual-isotope case, 74% for $^{171}$Yb-g with zoned measurement, 75% for $^{171}$Yb-g with shelving, and 78% for $^{174}$Yb clock qubits with zoned measurement — i.e. the 74–80% range quoted in the abstract.

• Halving Rydberg decay buys a factor of 1.7. Under reduced Rydberg decay, the error suppression factor would improve by roughly 1.7 if the decay rate were halved; the gain is larger still if depolarising errors are reduced simultaneously, even modestly.

• g-qubits beat m-qubits for data. $^{171}$Yb ground-state (g) encoding outperforms $^{171}$Yb metastable (m) encoding, attributed mainly to decay errors from the metastable state to the ground state (DECAY_mg channels). The paper notes these could in principle be converted into detectable erasure, but that leakage-detection gadgets add syndrome-extraction overhead, which is left to future work.

• SWAP syndrome extraction underperforms, contrary to expectation. Although prior studies suggested SWAP-based extraction would help by refreshing atoms, in these Yb simulations normal syndrome extraction gives lower logical error rates than SWAP extraction. The authors attribute this to the extra gate operations SWAP extraction requires, whose added errors outweigh the benefit of atom refreshment — at least when erasure information is not used in decoding.

• XZZX is comparable or better. The XZZX code matches the rotated surface code closely for the dual-isotope configurations and beats it for single-isotope ones; for $^{171}$Yb-g with shelving measurement, the per-round logical error rate is approximately half that of the rotated code at $d = 21$.

• Erasure flags were not used. The readout scheme can distinguish $|0\rangle$, $|1\rangle$, and loss, but the authors deliberately do not feed atom-loss flags into decoding, which they identify as a reason single-isotope alternatives may look worse here than they would with erasure-aware decoding.

Methodology in Plain English

The authors first write down how each kind of Yb qubit is encoded and how each operation — single-qubit gates, Rydberg two-qubit gates, clock-transition excitation, measurement, reset, idling, decay, and atom transport — can go wrong, including leakage into states outside the computational subspace. Because simulating these non-Pauli errors exactly is prohibitively expensive, they use a generalised Pauli twirling approximation that converts each noisy operation into two pieces: an ordinary Pauli error channel with certain probabilities, and an "erasure" channel representing population that leaves the qubit subspace and does not come back without a reset. This makes the whole circuit Clifford-simulatable.

They then build surface-code memory circuits — both the standard rotated surface code and the XZZX variant on the same lattice — with $r = d$ syndrome-extraction rounds, insert the twirled noise using DualYbSim, and sample $10^5$–$10^7$ shots per data point with stim. Syndromes are decoded with PyMatching using correlated matching, and the per-round logical error rate is extracted from the multi-round result.

For the error budget, they rescale each physical noise parameter individually by factors from 0.4 to 1.1 in steps of 0.025, measure $\Lambda^{-1}$ at $d = 5$ and $d = 7$, and fit a weighted quadratic model of $\Lambda^{-1}$ in the individual error rates. Quadratic terms are retained only for Rydberg-decay-related errors; any residual mismatch is recorded as "stray error."

Why This Matters

The paper converts an architectural intuition — "dual-species arrays let you measure in place, so they should be better" — into a quantitative, hardware-grounded claim, and in doing so separates which engineering improvements actually matter from which ones merely seem attractive. It also provides the community with reusable simulation infrastructure rather than a one-off result.

• Hardware roadmap prioritisation: the error budget points squarely at Rydberg-state lifetime and two-qubit gate quality as the highest-leverage targets, rather than, say, transport speed or readout.
• Architecture selection for neutral-atom QEC: groups deciding between single-isotope designs (with shelving or zoned transport) and dual-isotope designs get a directly comparable logical-error-rate curve.
• Code selection: the rotated-versus-XZZX comparison informs which decoder/code pairing to build first, particularly given XZZX's advantage in the noisier single-isotope regimes.
• Simulation tooling: DualYbSim lets other groups plug realistic Yb noise into stim circuits without rebuilding the twirling machinery.

Industry relevance: neutral-atom quantum computing is an active commercial and academic hardware race, and fault-tolerance projections depend heavily on assumed noise. A publicly released, experimentally motivated noise wrapper lowers the barrier for hardware teams to benchmark their own devices against these architectural baselines, and gives funders and system architects a clearer picture of which component specifications gate the transition to fault tolerance.

Future Directions

• Erasure-aware decoding. The authors repeatedly flag that the detected loss/leakage information — currently discarded — could substantially improve all configurations, and could change the ranking between dual-isotope in-place measurement and SWAP-based schemes.
• Leakage-detection gadgets for m-qubits. Converting metastable-to-ground decay into detectable erasure might rescue the m-qubit encoding, but the added syndrome-extraction overhead needs to be quantified.
• Co-optimising decay and depolarising errors. Since the suppression gain exceeds the naive expectation when both are reduced together, the paper motivates joint rather than isolated improvements to gate fidelity.
• Extending the error-budget fit. The quadratic approximation neglects cross-terms $p_i p_j$ and lumps unmodelled contributions into a "stray error" term, which the authors note grows near threshold or for distance-dependent noise; refining this is an open methodological question.

Target Audience

Researchers and engineers working on neutral-atom quantum computing and quantum error correction; architects choosing between single- and dual-species array designs; and graduate students or tool-builders who want a realistic, ready-to-use Yb noise model for surface-code simulations. Readers without background in stabiliser codes or Pauli twirling will find the results section accessible but the methods section demanding.

Authors’ abstract

Neutral-atom quantum computers are a promising platform for fault-tolerant quantum computation, but logical performance depends on systemic realistic noise factors during syndrome extraction. In dual-isotope Yb arrays, the roles of data and ancilla qubits are separated spectrally, allowing ancilla qubits to be measured in place without additional transport or shelving operations. Here we quantify the advantage of a dual-isotope Yb architecture for surface code memories. We develop an experimentally motivated Clifford-compatible noise model for dual-isotope 171Yb-174Yb systems using generalised Pauli twirling and implement it as a wrapper for Stim called DualYbSim, which has been packaged as an open source Python library. Simulations of rotated and XZZX surface codes show that a dual-isotope architecture with in-place measurement achieves the lowest logical error rates among the architectures considered, outperforming single-isotope schemes based on shelving or zoned measurement. Our error-budget analysis also identifies Rydberg-state decay as the dominant limitation, contributing to 74-80% of the logical error rate scaling, highlighting concrete experimental targets for improving FTQC performance.