Skip to content

Programmable Quantum Matter - Heralding Large Cluster States in Driven Inhomogeneous Spin Ensembles

FieldValue
JournalPRX Quantum
Published2026-05-06
AuthorsPratyush Anand, Louis Follet, Odiel Hooybergs, Dirk R. Englund
DOI10.1103/kb8s-sffj

The proposed SAFE-GRAPE framework reduces global control overhead from (O(N_q)) to (O(1)) for inhomogeneous spin ensembles of about 10 to 100 emitters. Its simulations retain single-qubit fidelity above 99.99% for simultaneous normalized pulse-length and frequency-detuning errors up to 0.3. Fidelity stays above 99% at errors of 0.4. Simulated decoupling yields ensemble-average coherence times more than a factor of seven longer than interleaved bang-bang CPMG.

The modeled network uses two spatially separated ensembles of negatively charged silicon-vacancy centers, written SiV−, in diamond. One global strain drive acts on every center in each ensemble. Local strain shifts spin transitions, while strain-dependent optical frequencies supply labels for later entanglement.

Conventional per-emitter interleaving requires control operations that scale as (O(N_q)). The composite waveform addresses the whole ensemble once. It must absorb spatial variation in strain amplitude, center-to-center differences in spin-strain susceptibility, and minor drive-power instability.

Each measurement cycle combines global initialization, global dynamical decoupling, pairwise single-photon entanglement, and compilation into a bipartite graph. The resulting bipartite cluster state is intended for measurement-based quantum computing. All performance values in the paper come from simulations and theoretical analysis, not a hardware demonstration.

Robust pulses under amplitude and detuning errors

Section titled “Robust pulses under amplitude and detuning errors”

The pulse search starts from reduced CinBB, which combines a CORPSE pulse inside a BB1 broadband composite pulse. The demonstrated sequence contains six elements. SAFE-GRAPE optimizes their durations and phases. The loss function is based on average gate infidelity over a discretized grid of amplitude and detuning errors. The implementation uses PyTorch and the L-BFGS optimizer.

For a π-X gate, SAFE-GRAPE reached 99.987% fidelity at simultaneous errors (\epsilon=f=0.25). Reduced CinBB reached 92.7% at the same point. SAFE-GRAPE caused a 51% increase in pulse duration and required more bandwidth than reduced CinBB. A bandwidth-aware version achieved similar infidelities with a 100 MHz I/Q drive.

The error map spans normalized amplitude and detuning errors from -0.3 to 0.3. At a nominal Rabi rate of (2\pi \times 30) MHz, (|f|\leq0.3) corresponds to (|\Delta|\leq2\pi\times9) MHz. This range targets residual drift after resonance identification and coarse tuning. It does not cover the full GHz-scale static frequency spread of an untreated SiV ensemble.

The authors evaluate filter functions numerically with the Python filter_function package. Sequence A uses global SAFE-GRAPE pulses, while sequence B uses interleaved bang-bang CPMG. As (f) changes from 0 to 0.3, A1’s low-frequency response rises from (4\times10^{-9}) to (1.8\times10^{-8}). B1 rises from (3\times10^{-9}) to (4.5\times10^{-7}). A1’s peak varies by less than 0.1%, while B1’s peak falls by 3%. The authors therefore judge A1 about 30 times more robust to detuning changes. Increasing A from one to 32 pulses lowers its low-frequency filter from about (10^{-5}) to below (10^{-12}).

The thermal model places the SiV− device on a 100 mK dilution-refrigerator stage. Sequence B interleaves 121 individually addressed CPMG sequences in the reported simulation. It creates more incoming pulses and a larger heat load than sequence A.

At a 0.2 ms decoupling window, B raises the modeled SiV− temperature by about 5.2%. A raises it by about 0.3%. Increasing the window from 0.2 ms to 2 ms lowers the pulse duty cycle and the temperature rise for both methods.

For a 0.2 ms window and (m=121), the reported grid averages are:

Pulse count NT2 for A (ms)T2 for B (ms)Mean A/B factor
20.43 ± 0.180.281.5
40.51 ± 0.270.192.7
80.49 ± 0.290.104.7
160.39 ± 0.190.067.2

At a 0.8 ms window, A reaches (0.65\pm0.47) ms at N=8, compared with 0.37 ms for B. At N=16, A reaches (0.62\pm0.44) ms, compared with 0.20 ms for B. B enters the region where (T_2<t_\mathrm{dds}), meaning it no longer provides useful decoupling. A remains partly outside that region.

Each method has an optimum N because more pulses improve filtering but also increase heating. The advantage is not universal at every point. At N=2 and a 0.8 ms window, B reaches 0.59 ms, compared with 0.45 ms for A. The B baseline excludes crosstalk, so it is an optimistic comparison.

Entanglement performance in finite windows

Section titled “Entanglement performance in finite windows”

The modified single-photon protocol starts from the initial ground state and creates heralded Bell pairs between the two systems. The simulations use (\alpha=10^{-4}) and (\eta=10^{-2}), with (\eta) defined as the single-photon detection efficiency. Each system contains 121 points in the error grid, interpreted as 121 physically distinct emitters. The two systems therefore have 14,641 possible links. The compilation window is 10 µs.

A link is counted when its time-averaged two-qubit error remains below 0.5. The effective error averages those usable links. Entanglement error is approximated as CNOT error, and the model adds local single-qubit error.

At a 0.2 ms decoupling window, B produces all 14,641 links at N=2. Its link count falls to zero for N above 2. A produces 10,231 links at N=2, 9,221 at N=4, 8,029 at N=8, and 5,857 at N=16. At 0.8 ms, B produces zero links for every tested N. A reaches 926 links at N=8, then falls to 447 links at N=16.

At N=16 and a 0.2 ms window, A has an effective error of 0.30, while B has 1.0. At N=8 and a 0.8 ms window, the values are 0.36 for A and 1.0 for B. The simulated link-count advantage ranges from order (10^2) to (10^4), depending on the decoupling window. The broader simulations were performed in Python with QuTiP and PyTorch.

Section titled “Compiling links into a bipartite cluster state”

Algorithm 1 constructs a global strain trajectory that moves each center’s optical transition past a common laser frequency. The mapping must be monotonic and injective so each crossing identifies one qubit. The authors simulate 11 SiV− centers with bias-strain standard deviation (6\times10^{-5}). Their common window contains one positive-slope crossing per center within (|\Delta f|<0.2) GHz.

Algorithm 2 overlaps the two photon timing trains and selects the ratio (m_\mathrm{scal}=T_A/T_B). It chooses this ratio to minimize the fraction of links that never occur. For a five-qubit system A and four-qubit system B with (m_\mathrm{scal}=1), every possible link is attempted after 20 (T_A). Simulations also find configurations with no missing links for system sizes up to 10. The two algorithms give a unique-link count on the order of (N_q). Sweeping (m_\mathrm{scal}) can improve that count.

The filter analysis assumes pure dephasing. It also gives every qubit the same noise spectrum and the same noise sensitivity. The thermal results depend on estimated normalization and assumed refrigerator, cable, sample, and anchoring parameters. The entanglement results use fixed (\alpha) and (\eta), and treat the CNOT error term as an approximation to the heralded link error.

The pulse robustness map covers residual detuning after selecting a narrow resonance class. It does not correct the full GHz-scale spread across an untreated ensemble. The longer SAFE-GRAPE pulse also imposes a bandwidth and duration cost. The authors identify charge-state instability as a practical hurdle for nanomechanical structures. They propose machine-learning-driven in situ optimization for environmental changes and improved mechanical mode shapes for strain transfer. The global drive, optimized pulses, thermal behavior, and compiled links were not demonstrated in hardware.

Engineering implications for SiV− diamond

Section titled “Engineering implications for SiV− diamond”

For this architecture, candidate diamond should provide optically active SiV− centers, a narrow residual spectral class, stable charge states, and a mechanical mode that couples strain to the addressed centers. The working-window simulation used transverse-oriented centers, so orientation and local strain susceptibility enter material and geometry selection.

Thermal design matters as much as pulse quality. The comparison shows that one global composite pulse train can reduce the heating associated with 121 interleaved control trains on a 100 mK stage. Control electronics should also support the bandwidth required by the composite sequence. The paper’s bandwidth-aware result used a 100 MHz I/Q model. Nearby nuclear spin states could serve as memory qubits during repeated entanglement attempts.

The paper needs diamond containing optically active SiV− centers for strain-tuned spin-photon interfaces, plus a nanomechanical structure that couples strain to the defects. 6CCVD’s diamond materials catalog lists MPCVD-grown single-crystal and polycrystalline plates and wafers up to 125 mm for base-material process development. The catalog does not specify SiV− creation or charge-state performance, so defect selection, strain coupling, and 100 mK integration should be reviewed through engineering support before production.

View Original Abstract

Atom-like emitters in solids have emerged as promising platforms for quantum sensing and information processing. Among the major challenges are inhomogeneities in emitter fine structure, which complicates quantum control. Here, we introduce a framework that leverages this emitter diversity to simplify the experimental resources needed to create optically heralded spin cluster states across <a:math xmlns:a=“http://www.w3.org/1998/Math/MathML” display=“inline”> <a:msub> <a:mi>N</a:mi> <a:mi>q</a:mi> </a:msub> </a:math> emitters from the conventional order <c:math xmlns:c=“http://www.w3.org/1998/Math/MathML” display=“inline”> <c:mi>O</c:mi> <c:mo stretchy=“false”>(</c:mo> <c:msub> <c:mi>N</c:mi> <c:mi>q</c:mi> </c:msub> <c:mo stretchy=“false”>)</c:mo> </c:math> to <g:math xmlns:g=“http://www.w3.org/1998/Math/MathML” display=“inline”> <g:mi>O</g:mi> <g:mo stretchy=“false”>(</g:mo> <g:mn>1</g:mn> <g:mo stretchy=“false”>)</g:mo> </g:math> within ensembles of <k:math xmlns:k=“http://www.w3.org/1998/Math/MathML” display=“inline”> <k:msub> <k:mi>N</k:mi> <k:mi>q</k:mi> </k:msub> <k:mo>∼</k:mo> <k:mn>10</k:mn> <k:mo>-</k:mo> <k:mn>100</k:mn> </k:math> . Specifically, the optimized pulse sequence simultaneously corrects parameter variations (pulse-length error and frequency detuning error), achieving single-qubit gate fidelities exceeding 99.99% for errors (normalized relative to Rabi-drive strength) up to 0.3, while maintaining fidelities above 99% even for errors as large as 0.4. Applying this optimized pulse sequence in the form of a Carr-Purcell-Meiboom-Gill (CPMG) based dynamical decoupling protocol to the dominant noise spectral density of silicon-vacancy centers in diamond, our approach enhances ensemble-average coherence times by more than a factor of <m:math xmlns:m=“http://www.w3.org/1998/Math/MathML” display=“inline”> <m:mn>7</m:mn> </m:math> relative to interleaved bang-bang-based CPMG. For state-of-the-art dilution refrigeration systems, we further estimate sharply reduced heating when driving a global resonant optimal dynamical decoupling across <o:math xmlns:o=“http://www.w3.org/1998/Math/MathML” display=“inline”> <o:msub> <o:mi>N</o:mi> <o:mi>q</o:mi> </o:msub> </o:math> silicon-vacancy spins, potentially resolving the current trade-off between spin coherence and scaling to <q:math xmlns:q=“http://www.w3.org/1998/Math/MathML” display=“inline”> <q:msub> <q:mi>N</q:mi> <q:mi>q</q:mi> </q:msub> <q:mo>≫</q:mo> <q:mn>1</q:mn> </q:math> . We further introduce a modified single-photon entanglement protocol with an efficient algorithm for deterministic entanglement compilation. Depending on the decoupling window, our method yields order <s:math xmlns:s=“http://www.w3.org/1998/Math/MathML” display=“inline”> <s:mi>O</s:mi> <s:mo stretchy=“false”>(</s:mo> <s:msup> <s:mn>10</s:mn> <s:mn>2</s:mn> </s:msup> <s:mo>-</s:mo> <s:msup> <s:mn>10</s:mn> <s:mn>4</s:mn> </s:msup> <s:mo stretchy=“false”>)</s:mo> </s:math> more entanglement links than bang-bang sequences, with theoretical guarantees of order <w:math xmlns:w=“http://www.w3.org/1998/Math/MathML” display=“inline”> <w:mi mathvariant=“normal”>Ω</w:mi> <w:mo stretchy=“false”>(</w:mo> <w:msub> <w:mi>N</w:mi> <w:mi>q</w:mi> </w:msub> <w:mo stretchy=“false”>)</w:mo> </w:math> unique links—improvable via control-parameter tuning. Our approach thus offers enhanced fidelity, scalability, and robustness. Together, these techniques provide foundational tools—including global unitary control, phase denoising, remote entanglement, and compilation—for scalable quantum computing architectures based on heterogeneous spin ensembles.