Abstract
As superconducting quantum circuits scale to larger sizes, the problem of frequency crowding proves a formidable task. Here we present a solution for this problem in fixedfrequency qubit architectures. By systematically adjusting qubit frequencies postfabrication, we show a nearly tenfold improvement in the precision of setting qubit frequencies. To assess scalability, we identify the types of “frequency collisions” that will impair a transmon qubit and crossresonance gate architecture. Using statistical modeling, we compute the probability of evading all such conditions, as a function of qubit frequency precision. We find that, without postfabrication tuning, the probability of finding a workable lattice quickly approaches 0. However, with the demonstrated precisions it is possible to find collisionfree lattices with favorable yield. These techniques and models are currently employed in available quantum systems and will be indispensable as systems continue to scale to larger sizes.
Introduction
Realizing robust largescale quantum information processors is one of the foremost challenges in quantum science. Many practical applications have been proposed for robust quantum computers, including estimating the ground state energy of chemical compounds and implementing machine learning algorithms^{1,2,3,4,5,6,7,8}. Quantum advantage relative to classical computers can be realized without full fault tolerance, but requires large quantum circuits that a classical computer cannot simulate^{9}. Recent demonstrations have shown qubit circuits nearly at the threshold for demonstrating quantum advantage^{10}. Much work remains in order to realize faulttolerant quantum processors; however, scaleup of solidstate quantum circuits has shown consistent and ongoing progress^{11,12,13,14,15,16,17,18,19,20}. As the qubit circuits are scaled up, they must maintain high one and twoqubit gate fidelities, high qubit connectivity, and low crosstalk error, which can be measured in a holistic sense via the quantum volume of the circuit^{21,22}. Lattices of fixedfrequency transmon qubits represent a promising architecture for building systems of larger sizes^{10}. A growing number of systems at the 20–50qubit scale are now available to users through cloud access. Fixedfrequency transmons are largely insensitive to charge or flux noise and have achieved coherence times of 100 μs and growing. A variety of technical challenges confront further system scaling, including improving threedimensional circuit integration and fast readout. High on the list of such challenges is the issue of “frequency crowding.”
The crossresistance (CR) gate, a hardwareefficient allmicrowave gate^{23,24,25,26}, is readily used to entangle fixedfrequency transmons with gate fidelities >99%, approaching the threshold for faulttolerant codes^{27}. To achieve these fidelities, the CR gate needs not only high coherence qubits but also a precise setting of the qubits’ frequencies. The CR gate activates a ZX interaction by driving one “control” qubit with a microwave pulse at the other “target” qubit’s transition frequency. The magnitude of the ZX as well as other Hamiltonian terms depends on the relative frequencies of the two qubits^{28,29}. Diminished ZX magnitude increases gate time, while other terms such as ZZ add gate errors. Neighboring qubits having the wrong detuning will exhibit a frequency collision in which the ZX may be suppressed or other undesirable effects arise.
Maintaining high gate fidelities for all pairs in a lattice will require solving this frequencycrowding problem by precise setting of qubit frequencies to specified values, as characterized by a standard deviation σ_{f}. To achieve low σ_{f}, the tunneljunction conductance must be controlled with high precision. Transmon frequency f_{01} follows \(h{f}_{01}\simeq \sqrt{8{E}_{J}{E}_{C}}{E}_{C}\), where Josephson energy \({E}_{J}=\frac{\hslash {I}_{{{{\rm{c}}}}}}{2e}\) is many times greater than charging energy \({E}_{C}=\frac{{e}^{2}}{2C}\)^{30}. In typical transmons, a photolithographically defined capacitance C has dimensions in the tens to hundreds of microns and varies little from qubit to qubit. The critical current I_{c} is set by a tunnel barrier of area ~100 × 100 nm and thickness a few nm and is thus challenging to fabricate with precision better than a few percent^{31,32,33,34,35}. However, tunnel barrier resistance R_{n} is readily measurable to precision better than 0.1% and relates to I_{c} according to the Ambegaokar–Baratoff relation \({I}_{{{{\rm{c}}}}}=\frac{\pi {{\Delta }}}{2e{R}_{{{{\rm{n}}}}}}\) (where Δ is the superconducting gap energy)^{36}. We expect imprecision in resistance σ_{R} to produce a corresponding imprecision in frequency \({\sigma }_{f}=\frac{1}{2}\frac{{\sigma }_{R}}{\left\langle R\right\rangle }\cdot \left\langle f\right\rangle\), where \(\left\langle f\right\rangle\) and \(\left\langle R\right\rangle\) are the mean values of frequency and resistance, respectively. We can therefore measure R_{n} before a chip is cooled in order to assess qubit frequency imprecision. The best demonstrated precision in setting R_{n} at the time of fabrication is 2%^{34}. A 2% variation in R_{n} indicates a fractional σ_{f} of 1%.
Careful design of lattices can enable error correction codes while at the same time minimizing the likelihood of “frequency collisions” and therefore the required σ_{f} for fabrication yield^{37,38}. Yet even the most robust designs require a fractional σ_{f} of 0.25–0.5%, which represents a factor of 2–4 improvement over the best literature results. To overcome such limits will require rework of individual qubits’ tunnel junctions after fabrication. Thermal anneal has been shown to increase tunnel resistance R_{n}, and laser heating has been demonstrated as a highly localized rework tool^{39,40,41,42,43,44}. However, the inherent variability of the anneal process itself must be overcome, and qubit frequency control utilizing such techniques at scale has never been presented in the literature.
In this paper, we introduce an adaptive postfabrication trimming technique that we use to incrementally adjust R_{n} on a qubitbyqubit basis, thereby overcoming inherent variability in both initial qubit fabrication and the laser anneal. We demonstrate this improvement in qubit frequency precision clearly in terms of narrowed frequency distributions. Crucially, we demonstrate qubit frequency imprecision σ_{f} of the same magnitude as the imprecision of predicting f_{01} from R_{n}. To estimate the scalability of this technique for the fabrication of errorcorrected lattices, we employ a statistical yield model based on σ_{f} relative to specific collision bounds. This model predicts the severity of the frequencycrowding problem for different topologies and scales of errorcorrected multiqubit lattices as a function of code distance. The model demonstrates that, using conventional transmon fabrication, scaledup qubit lattices will fail to evade frequency collisions. In contrast, our trimming technique achieves adequate σ_{f} for scalable fabrication of distance3 through distance7 heavysquare and heavyhexagon codes. In particular, this technique enables the high yield fabrication of the distance3 and distance5 heavyhexagon lattices currently deployed as IBM cloud connected systems^{22}.
Results
Frequency precision σ _{f} from transmon fabrication
To assess the σ_{f} resulting from qubit fabrication, we developed a test vehicle containing a large number of identically fabricated qubits (Fig. 1). We cooled the chip in a dilution refrigerator and used dispersive readout through halfwave microwave resonators to measure qubit frequencies^{45}. We measured the frequencies of 31 qubits to a precision better than 100 kHz using a Ramsey fringe method. The qubit frequencies had random variation σ_{f} = 132.3 MHz (Fig. 1) or 2.3% of the median frequency. After warming the qubits to room temperature, we measured their junction resistances. The standard deviation σ_{R} was 365 Ω, 4.6% of the median R_{n}. Fractional σ_{f} is exactly half of fractional σ_{R}, as expected from Transmon theory and the Ambegaokar–Baratoff relation. A plot of R_{n} against transmon frequency (Fig. 1) fits a power law of approximately \(\frac{1}{2}\) power, as expected from theory. To further assess the fidelity of the frequencies to this fvsR_{n} correlation, we show the residual scatter after subtracting the fit line. This appears in the inset in Fig. 1 and exhibits a standard deviation 14.5 MHz or 0.25% of the qubit median frequency. Following transmon theory and the Ambegaokar–Baratoff relation, this residual scatter could indicate a qubittoqubit variation of up to 0.5% in superconducting gap Δ or qubit capacitance C. Small systematic errors in measuring R_{n}, for instance, due to substrate conductance, could also contribute. As we discuss below, future scaling of superconducting quantum logic circuits will require improvements in σ_{f} and therefore better control of these parameters.
Tuning using selective laser anneal
To reduce σ_{f}, we developed a technique for selective laser anneal to shift tunnel resistance R_{n} by precalibrated increments (see “Methods” and Fig. 2). We demonstrate the achievable frequency control of this technique by shifting the 31 measured qubits into a twofrequency pattern. We employed an R_{n} vs f correlation (Fig. 1) to designate the target resistances. We shifted 16 junctions to one R_{n} group and 15 to another R_{n} group. After tuning, the group of 16 junctions had median resistance 7.984 kΩ and the group of 15 had median resistance 8.798 kΩ. The 31 junctions clustered around these medians with an overall precision of σ_{R} = 51 Ω, about 0.61%. In a dilution refrigerator, we remeasured the frequencies of the qubits in the two groups. A change in fridge instrumentation degraded the readout signaltonoise ratio of two of the qubits, so that their remeasurement (using continuouswave spectroscopy) had frequency precision of only 2 MHz. The others were remeasured in the same way as in the first cooldown. The resulting frequencies appear in Fig. 1. The two frequency groups are approximately normally distributed and have medians f_{0,1} = 5.430 GHz and f_{0,2} = 5.7046 GHz. Calculating \({\sigma }_{f}=\sqrt{\left\langle {({f}_{i}{f}_{0,j})}^{2}\right\rangle }\), where f_{0,j} represents f_{0,1} or f_{0,2} as appropriate for a given qubit Q_{i}, we assess the overall precision σ_{f} = 14.0 MHz. This imprecision is nearly identical to the residual scatter from the f(R) fit line (Fig. 1), which guided the tuning, and the fractional precision \({\sigma }_{f}/\left\langle f\right\rangle =\) 0.25% is slightly better than half of the fractional precision in setting R_{n}. Drift in R_{n} reported in the literature^{39} does not appear to be a limiting factor in this study. As we show in “Methods,” the laseranneal tuning technique is capable of precisions of 0.3% in R_{n}. In future work, the imprecision of 14.5 MHz in predicting f from R_{n} could be made the limiting imprecision in σ_{f}.
Discussion
Our postfabrication trimming reduced σ_{f} by 9.5× compared to initial fabrication. To assess whether this level of precision is sufficient to reliably prepare lattices of fixedfrequency transmons capable of errorcorrecting codes, we must quantify the frequencycrowding problem. Transmon qubits are weakly anharmonic and have decreasing transition energies at higher levels. Therefore, degeneracies among the \(\left0\right\rangle \to \left1\right\rangle\), \(\left1\right\rangle \to \left2\right\rangle\), and \(\left0\right\rangle \to \left2\right\rangle\) transitions of nearby qubits can all contribute to frequency collisions. We must consider the relative frequencies of both nearest neighbors and next nearest neighbors in the lattice^{28,46,47}. Figure 3 illustrates the relative positions of nearestneighbor and nextnearestneighbor qubits in a section of lattice, and Table 1 lists the seven cases most likely to lead to gate errors^{28}. We can think of them qualitatively as follows: Type 1 causes hybridization of states in Q_{j} and Q_{k}, while in type 2 the CR pulse excites Q_{j} into the noncomputational \(\left2\right\rangle\) state. Type 3 excites Q_{k} to the \(\left2\right\rangle\) state but does not require a CR tone. In condition 4, ZX is weak, which implies long gate times and increased gate error^{28,29}. In type 5, the CR gate addresses an additional neighboring qubit. In type 6, when one qubit is the target of a CR gate, its next nearest neighbor leaks to the \(\left2\right\rangle\) state. In type 7, Q_{j} acts as the control, Q_{i} or Q_{k} as the target, and the third qubit constitutes a “spectator.” An excitedstate spectator can emit a photon that combines with photons in the CR pulse to excite Q_{j} into the \(\left2\right\rangle\) state.
Around each of the frequency collisions described in Table 1, we can designate a window of undesired frequencies. This breaks the frequency space into allowed and forbidden regions. Type 4 listed in Table 1 defines forbidden zones where ZX coupling is too low. For the other six conditions, we forbid regions where the frequency collision is the dominant source of gate error. Existing multiqubit systems with CR gates typically exhibit twoqubit gate errors of 1–2% regardless of frequency^{15,46}. Reference ^{28} considers an effective Hamiltonian model for the CR gate, as a function of the relative frequency of control and target qubits. From this model, we estimate the frequency windows for nearestneighbor collisions (Table 1, types 1–3). Within these windows, assuming typical gate parameters of 30–50 MHz drive amplitude and 200–400 ns duration, we expect errors exceeding ~1%. Here we make an assumption that similar bounds apply to nextnearestneighbor interactions (types 5–7). Future lattice scaling will benefit from defining frequency collisions precisely to achieve specific quantum volume or errorcorrection thresholds. Such work can exploit numerical models^{22,48} that find CR gate error as a function of frequency, coupling, and rate.
A useful lattice of qubits should enable high quantum volume and faulttolerant operation while avoiding all of the frequency collisions and forbidden regions presented in Table 1. Both lattice layout and the pattern of qubit frequencies are relevant. We consider three types of lattices: square, “heavy square” and “heavy hexagon” (Fig. 3). Lattices comprise qubits and twoqubit connections, each qubit being linked to no more than four neighbors. In many practical implementations, these links comprise microwaveresonant buses. A square lattice facilitates “surface code” faulttolerant codes^{49}. Recent literature describes hybrids of the surface code with Bacon–Shortype codes, which can be employed in heavyhexagon and heavysquare lattices to achieve fault tolerance, albeit with lower error thresholds than the surface code^{37}. In addition to the data and ancilla qubit roles employed in the surface code, these hybrid codes assign a portion of the lattice as “flag” qubits.
In the square lattice, every qubit in the bulk of the lattice lies on a degreefour vertex, while some at edges have degree two or degree three. If we populate the square lattice with five distinct frequencies of qubits, f_{5} > f_{4} > f_{3} > f_{2} > f_{1}, with appropriate spacing between the frequencies, we can avoid all the forbidden regions of Table 1^{13}. In Fig. 3, we illustrate this pattern for a square lattice capable of a distance5 (d = 5) rotated surface code. Condition 4 of Table 1 requires f_{control} > f_{target}, so the pattern also fixes the direction of CNOT gate for each pair.
In contrast to the square lattice, the heavysquare lattice includes both degreetwo and degreefour vertices in the bulk. Degreeone, degreetwo, or degreethree vertices appear at the edges. We take advantage of this pattern to make all the degreetwo vertices control qubits, using a threefrequency pattern f_{3} > f_{2} > f_{1}. Since every control qubit (frequency f_{3}) is linked to at most two target qubits, we need only two properly chosen targetqubit frequencies (f_{1} and f_{2}) to satisfy conditions 5–7 of Table 1, as shown in Fig. 3. A third type of lattice, the heavy hexagon, uses a similar scheme. Here the bulk of the lattice includes degreethree and degreetwo vertices. Additional degreetwo and degreeone vertices lie at the edges. In this lattice, all of the frequency collisions and forbidden regions can be satisfied using only three frequencies f_{3} > f_{2} > f_{1}, with all control qubits residing on degreetwo vertices with frequency f_{3}.
We use a Monte Carlo model to quantify the frequency crowding in each lattice type. We sample the qubits at random frequencies drawn from normal distributions characterized by σ_{f} and count the collisions defined in Table 1 (see “Methods” and Fig. 6). In Fig. 4, we show the mean number of frequency collisions predicted by the Monte Carlo model for each lattice type and frequency pattern, as a function of σ_{f}. As σ_{f} → 0, the lattice approaches the ideal patterns of Fig. 3 and has zero frequency collisions. As σ_{f} increases, the number of frequency collisions rises steadily. As σ_{f} → f_{01} − f_{12}, the different conditions appearing in Table 1 all become likely, and a limiting number of frequency collisions is reached. Yield follows the inverse trend, as seen in Fig. 4. As σ_{f} increases, the likelihood of finding a “collisionfree” chip falls off sharply. While the step sizes between frequency setpoints f_{1} to f_{5} are important, absolute values of setpoints are not. Setting f_{1} = 5.0, f_{2} = 5.07, and f_{3} = 5.14 GHz works as well as f_{1} = 5.05, f_{2} = 5.12, and f_{3} = 5.19 GHz.
The yield and mean collision number are a function of the several different collision types and bounds, so they are not readily susceptible to an analytic formulation. However, we can propose a simplified model for yield: in order for a lattice to be collisionfree, every qubit in the lattice must fall within some frequency “window” ±Δf relative to its setpoint. Presuming the qubit frequencies are normally distributed, the probability of this occurring goes as the cumulative distribution function, raised to the power N, where N is the number of qubits: \({\left[\int\nolimits_{\infty }^{({{\Delta }}f/{\sigma }_{f})}{e}^{\frac{1}{2}{x}^{2}}{\rm{d}}x\right]}^{N}\). In the yield plot in Fig. 4, we fit this expression to find Δf for each lattice.
These model results allow us to predict how different lattice types and frequency patterns will respond to fabrication imprecision. As shown in Fig. 4, if imprecision σ_{f} is >30 MHz, any d = 5 lattice will exhibit >10 frequency collisions of one or another of the types listed in Table 1, causing the affected gates to have error rates above ~1%. However, if σ_{f} = 10 MHz then on average the d = 5 square lattice will exhibit 5 frequency collisions, while the heavysquare and heavyhexagon lattices will exhibit 0.1 frequency collision. Considered in terms of yield, we see from Fig. 4 that if σ_{f} = 10 MHz, then for a d = 5 device, a square lattice with 5frequency pattern has a 0.8% likelihood to be collision free, whereas a heavysquare lattice with 3frequency pattern has 90% likelihood and heavy hexagon with 3frequency pattern has 92% likelihood. Alternatively we can ask, how well do we have to control σ_{f}? If we seek a 10% yield, then Fig. 4 indicates that, for a d = 5 device, a square lattice with 5frequency pattern requires σ_{f} < 8 MHz, whereas a heavysquare lattice with 3frequency pattern requires σ_{f} = 16 MHz and heavy hexagon with 3frequency pattern requires σ_{f} = 17 MHz. Although the square lattice requires 10–20% fewer qubits than the other types at each distance d, it requires far better frequency precision.
The asfabricated σ_{f} seen in Fig. 1 is 132.3 MHz (see “Results”). The Monte Carlo modeling finds that for a heavyhexagon lattice at d = 3 scale this σ_{f} can enable 0.1% yield of collisionfree chips. Other lattice types and larger scales will all have yield ≪0.1%. The retuned σ_{f} = 14.0 MHz demonstrated in Fig. 1 will improve the yield in all types of lattice. Predictions of the Monte Carlo model for σ_{f} = 14.0 MHz appear in Table 2. At d = 5 scale, the heavyhexagon and heavysquare lattices and 3frequency patterns should be collisionfree nearly onethird of the time, while at d = 7 scale the yield is about four times smaller, still reasonable for prototype systems.
As seen from the Monte Carlo analysis, the laseranneal rework method can scale to the >100 qubit size, enabling a wellchosen lattice and frequency pattern to implement d = 7 errorcorrection codes free of frequency crowding. To examine needs for the next generation of chips up to the 1000qubit level, we can coarsely estimate requirements by extrapolating the fixed window model for the heavyhexagon lattice as shown in Fig. 5. While the σ_{f} = 14.0 MHz demonstrated here enables practical yield up to the 100–200 qubit scale, it is clear that roughly a factor of two further improvement is needed to scale toward 1000 qubits. Since this precision is also better than the resistancetofrequency prediction precision shown in this work, development of further refinements in tuning and frequency prediction approaches will be necessary as the scale of fixedfrequency transmon circuits surpass the 100 qubit milestone.
Methods
Chip fabrication
A chip of the kind used to determine σ_{f} and to test our laseranneal rework process appears in Fig. 1. All microwave elements comprise Nb films ~200 nm thick on a silicon substrate. Each qubit is coupled to a readout resonator but is not directly coupled to any nearby qubits. All transmon capacitors are identical. Junctions are fabricated using identical electronbeam lithographic patterns and deposited simultaneously using doubleangle deposition and oxidation^{50}. The individual qubit design is similar to that used in ref. ^{27} with anharmonicity f_{12} − f_{01} ≃ −330 MHz. Junctions have linear dimension ~100 nm and are designed for I_{c} of ~30 nA. During packaging, we accidentally damaged 3 of the 36 qubits and found these to be nonfunctional when cooled in a dilution refrigerator. We left 2 of the remaining 33 qubits untuned as experimental controls, so that our tuning demonstration includes 31 qubits.
Tuning using selective laser anneal
We have built an integrated junction rework system that can measure and modify the junction resistance. Figure 2 shows a schematic of our laser annealing system, which we call Laser Annealing of Stochastically Impaired Qubits (LASIQ). The laser output is generated by a diodepumped solidstate laser, frequency doubled to 532 nm. Active power control of the anneal beam from approximately 1.7 to 2 W is performed using a piezorotary mounted waveplate and polarizing beam splitter, which is adaptively adjusted based on a pickoff beam measured on a downstream silicon photodiode. A precisiontimed shutter exposes the device for 0.3–10 s, and beam alignment is performed using a mechanical mirror mount, which directs the beam via pattern recognition to the transmon junction center. To achieve a more consistent anneal, the beam is shaped into a fourspot pattern, which avoids directly illuminating the junction but uniformly heats the surrounding substrate^{51}.
By careful control of laser power and pulse duration, we use this system to adjust R_{n}. This process overcomes the imprecision due to transmon fabrication, with a residual imprecision σ_{f} due to the rework process. To develop the process, we prepared a set of 126 junctions identical to qubit junctions and measured their response to a range of laser powers and exposure times. We recorded R_{n} shifts up to 15% relative to initial R_{n} for total anneal durations varying from 2 to 80 s and laser powers varying from 1.6 to 2 W. Response to laser power in particular was highly nonlinear. Based on these empirical calibrations of R_{n} shift to power and exposure, we established a qubit tuning process: We first measure the transmon junction’s R_{n} using fourpoint probing of the transmon capacitor pads at 25 °C. Using a f(R_{n}) prediction based on a previously determined correlation curve (Fig. 1), we assign the junction a target resistance corresponding to the target frequency in a multiqubit chip lattice. Because the anneal can shift R_{n} in only one direction, the target must be higher than the initial R_{n}. We anneal the qubit junction using laser power and duration chosen from our calibration set, then remeasure its R_{n}. By remeasuring after each anneal, we can adjust for random variation in the amount of R_{n} shift. A junction requiring large shifts in R_{n} may require repeated anneals to reach its target, as shown in Fig. 2. The control algorithm increases the resistance until the measured value is within 0.3% of the target value. In a separate trial of tuning precision, >300 junctions were tuned to target R_{n}s ranging from 0.4 to 14.5% above their initial values and landed successfully within this 0.3% margin. We observed this precision to be independent of the target R_{n}. We expect 0.3% imprecision in R_{n} to introduce 0.15% imprecision in transmon frequency.
Monte Carlo frequencycrowding model
Using a Monte Carlo model, we can estimate the incidence of frequency collisions in a lattice as a function of σ_{f}. We assume that imperfect frequency setting will distribute qubit frequencies normally around their design frequencies with standard deviation σ_{f}. For lattices of the type shown in Fig. 3, we designate 3–5 frequencies f_{1}, f_{2}, f_{3}, f_{4}, f_{5} spaced at regular intervals in the pattern shown. We set f_{1} = 5 GHz, similar to realworld transmons^{22,52}. We sample the qubit frequencies randomly around these values and count the collisions throughout the lattice, as listed in Table 1. To avoid unnecessary counting of type 4, we designate the higherfrequency qubit of every pair to be the control for that gate pair. This process is illustrated in Fig. 6. We repeat the frequency assignment and counting to build statistics for a given lattice and frequency pattern. We then repeat the model for a range of σ_{f} values from 0 to 150 MHz. We repeat the entire process over a range of frequency spacings to find the spacing that minimizes frequency collisions at each value of σ_{f}. As a function of σ_{f}, we can then extract (1) the mean number of total collisions in the lattice and (2) the fraction of repetitions that result in zero collisions (yield). Our simulations used 1000 repetitions except to find yield <1% in d = 5 lattices and <0.2% in d = 3 lattices, which used 4000 repetitions, and in d = 7 lattices to find mean collisions for σ_{f} < 16 MHz or yield >50% (100 repetitions) or to find mean collisions for σ_{f} > 16 MHz (40 repetitions).
Data availability
The experimental data presented in this manuscript are available from the corresponding author upon reasonable request.
References
Kandala, A. et al. Hardwareefficient variational quantum eigensolver for small molecules and quantum magnets. Nature 549, 242–246 (2017).
Havlíček, V. et al. Supervised learning with quantumenhanced feature spaces. Nature 567, 209–212 (2019).
Hempel, C. et al. Quantum chemistry calculations on a trappedion quantum simulator. Phys. Rev. X 8, 031022 (2018).
Colless, J. et al. Computation of molecular spectra on a quantum processor with an errorresilient algorithm. Phys. Rev. X 8, 011021 (2018).
Nam, Y. et al. Groundstate energy estimation of the water molecule on a trappedion quantum computer. npj Quantum Inf. 6, 1–6 (2020).
Arute, F. et al. HartreeFock on a superconducting qubit quantum computer. Science 369, 1084–1089 (2020).
ArgüelloLuengo, J., GonzálezTudela, A., Shi, T., Zoller, P. & Cirac, J. I. Analogue quantum chemistry simulation. Nature 574, 215–218 (2019).
McArdle, S., Endo, S., AspuruGuzik, A., Benjamin, S. C. & Yuan, X. Quantum computational chemistry. Rev. Mod. Phys. 92, 015003 (2020).
Kandala, A. et al. Error mitigation extends the computational reach of a noisy quantum processor. Nature 567, 491–495 (2019).
Córcoles, A. D. et al. Challenges and opportunities of nearterm quantum computing systems. Proc. IEEE 1–15 https://doi.org/10.1109/JPROC.2019.2954005 (2019).
Chow, J. M. et al. Implementing a strand of a scalable faulttolerant quantum computing fabric. Nat. Commun. 5, 4015 (2014).
Córcoles, A. et al. Demonstration of a quantum error detection code using a square lattice of four superconducting qubits. Nat. Commun. 6, 6979 (2015).
Gambetta, J. M., Chow, J. M. & Steffen, M. Building logical qubits in a superconducting quantum computing system. npj Quantum Inf. 3, 2 (2017).
Takita, M. et al. Demonstration of weightfour parity measurements in the surface code architecture. Phys. Rev. Lett. 117, 210505 (2016).
Takita, M., Cross, A. W., Córcoles, A. D., Chow, J. M. & Gambetta, J. M. Experimental demonstration of faulttolerant state preparation with superconducting qubits. Phys. Rev. Lett. 119, 180501 (2017).
Kelly, J. et al. State preservation by repetitive error detection in a superconducting quantum circuit. Nature 519, 66–69 (2015).
Risté, D. et al. Detecting bitflip errors in a logical qubit using stabilizer measurements. Nat. Commun. 6, 6983 (2015).
Ofek, N. et al. Extending the lifetime of a quantum bit with error correction in superconducting circuits. Nature 536, 441–445 (2016).
Arute, F. et al. Quantum supremacy using a programmable superconducting processor. Nature 574, 505–510 (2019).
Kjaergaard, M. et al. Superconducting qubits: current state of play. Annu. Rev. Condens. Matter Phys. 11, 369–395 (2020).
Cross, A. W., Bishop, L. S., Sheldon, S., Nation, P. D. & Gambetta, J. M. Validating quantum computers using randomized model circuits. Phys. Rev. A 100, 032328 (2019).
Jurcevic, P. et al. Demonstration of quantum volume 64 on a superconducting quantum computing system. Preprint at https://doi.org/10.1088/20589565/abe519 (2020).
Chow, J. M. et al. Simple allmicrowave entangling gate for fixedfrequency superconducting qubits. Phys. Rev. Lett. 107, 080502 (2011).
Chow, J. M. et al. Universal quantum gate set approaching faulttolerant thresholds with superconducting qubits. Phys. Rev. Lett. 109, 060501 (2012).
de Groot, P. C. et al. Selective darkening of degenerate transitions for implementing quantum controlledNOT gates. N. J. Phys. 14, 073038 (2012).
Rigetti, C. & Devoret, M. Fully microwavetunable universal gates in superconducting qubits with linear couplings and fixed transition frequencies. Phys. Rev. B 81, 134507 (2010).
Sheldon, S., Magesan, E., Chow, J. M. & Gambetta, J. M. Procedure for systematically tuning up crosstalk in the cross resonance gate. Phys. Rev. A 93, 060302 (2016).
Magesan, E. & Gambetta, J. M. Effective Hamiltonian models of the crossresonance gate. Phys. Rev. A 101, 052308 (2020).
Ware, M. et al. Crossresonance interactions between superconducting qubits with variable detuning. Preprint at http://arxiv.org/abs/1905.11480v1 (2019).
Koch, J. et al. Chargeinsensitive qubit design derived from the Cooper pair box. Phys. Rev. A 76, 042319 (2007).
Potts, A., Parker, G., Baumberg, J. & de Groot, P. CMOS compatible fabrication methods for submicron Josephson junction qubits. IEE Proc. Sci. Meas. Technol. 148, 225–228 (2001).
Wu, X. et al. Overlap junctions for high coherence superconducting qubits. Appl. Phys. Lett. 111, 032602 (2017).
Costache, M. V., Bridoux, G., Neumann, I. & Valenzuela, S. O. Lateral metallic devices made by a multiangle shadow evaporation technique. J. Vac. Sci. Technol. B 30, 04E105 (2012).
Kreikebaum, J. M., O’Brien, K. P., Morvan, A. & Siddiqi, I. Improving waferscale Josephson junction resistance variation in superconducting quantum coherent circuits. Supercond. Sci. Technol. 33, 06LT02 (2020).
Foroozani, N. et al. Development of transmon qubits solely from optical lithography on 300 mm wafers. Quantum Sci. Technol. 4, 025012 (2019).
Ambegaokar, V. & Baratoff, A. Tunneling between superconductors. Phys. Rev. Lett. 10, 486–489 (1963).
Chamberland, C., Zhu, G., Yoder, T. J., Hertzberg, J. B. & Cross, A. W. Topological and subsystem codes on lowdegree graphs with flag qubits. Phys. Rev. X 10, 011022 (2020).
Chow, J. M., Magesan, E., Steffen, M., Gambetta, J. M. & Takita, M. Reducing qubit frequency collisions through lattice design. US patent (10,622,536) US20200161529A1 (2020).
Koppinen, P. J., Väistö, L. M. & Maasilta, I. J. Complete stabilization and improvement of the characteristics of tunnel junctions by thermal annealing. Appl. Phys. Lett. 90, 053503 (2007).
Granata, C. et al. Trimming of critical current in niobium Josephson devices by laser annealing. J. Phys. Conf. Ser. 97, 012110 (2008).
Muthusubramanian, N. et al. Local trimming of transmon qubit frequency by laser annealing of Josephson junctions. In American Physical Society March Meeting B29.015 (American Physical Society, 2019).
Oliva, A. & Monaco, R. Annealing properties of high quality Nb/AlAlO_{x}/Nb tunnel junctions. IEEE Trans. Appl. Supercond. 4, 25–32 (1994).
Lehnert, T., Billon, D., Grassl, C. & Gundlach, K. H. Thermal annealing properties of NbAl/AlO_{x}Nb tunnel junctions. J. Appl. Phys. 72, 3165–3168 (1992).
Rosenblatt, S., Orcutt, J. S. & Chow, J. M. Laser annealing qubits for optimized frequency allocation. US patent 10340438B2 (2019).
Blais, A., Huang, R.S., Wallraff, A., Girvin, S. M. & Schoelkopf, R. J. Cavity quantum electrodynamics for superconducting electrical circuits: an architecture for quantum computation. Phys. Rev. A 69, 062320 (2004).
McKay, D. C., Sheldon, S., Smolin, J. A., Chow, J. M. & Gambetta, J. M. Threequbit randomized benchmarking. Phys. Rev. Lett. 122, 200502 (2019).
Malekakhlagh, M., Magesan, E. & McKay, D. C. Firstprinciples analysis of crossresonance gate operation. Phys. Rev. A 102, 042605 (2020).
Sundaresan, N. et al. Reducing unitary and spectator errors in cross resonance with optimized rotary echoes. PRX Quantum 1, 020318 (2020).
Fowler, A. G., Mariantoni, M., Martinis, J. M. & Cleland, A. N. Surface codes: towards practical largescale quantum computation. Phys. Rev. A 86, 032324 (2012).
Dolan, G. J. Offset masks for liftoff photoprocessing. Appl. Phys. Lett. 31, 337–339 (1977).
Rosenblatt, S. & Orcutt, J. S. Laser annealing of qubits with structured illumination. US patent 10170681B1 (2019).
Sheldon, S. et al. Characterizing errors on qubit operations via iterative randomized benchmarking. Phys. Rev. A 93, 012301 (2016).
Orcutt, J. S. Laser annealing of qubits using a diffractive beam splitter. US patent application (application pending) (2021).
Acknowledgements
We acknowledge funding from the Intelligence Advanced Research Projects Activity (IARPA) under contract W911NF1610114, for the multiqubit test vehicle and frequencyvsresistance correlation studies. We thank N. Bronn, M. Carroll, C. Chamberland, A. Cross, J. Gambetta, J. Ku, M. Malekakhlagh, D. McKay, B. Plourde, E. Pritchett, A. Rosenbluth, M. Takita, J. Timmerwilke, and G. Zhu for helpful discussions. We thank E. Porter for coding assistance, Y. Martin and R. Haight for assistance in constructing the laser optics, and R. Patel for photomicroscopy.
Author information
Authors and Affiliations
Contributions
J.B.H. developed the Monte Carlo model with input from J.A.S. and along with J.S.O. analyzed its results. E.M. developed theory to define frequency collisions. J.B.H., J.S.O., and J.M.C. conceived the experiment. V.P.A. and J.B.H. designed the device, and V.P.A. and M.B. fabricated it. M.S. developed apparatus and procedures used for qubit measurement. J.B.H. and M.S. measured the qubit frequencies and J.B.H. analyzed the data. J.B.Y. and E.J.Z. measured and analyzed roomtemperature resistance data. J.S.O., S.R., and E.J.Z. conceived and developed the LASIQ tuning method. J.S.O. and E.J.Z. built the LASIQ apparatus. E.J.Z. undertook the tuning experiments. J.B.H, J.S.O., and E.J.Z. prepared the manuscript, with input from all authors.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Hertzberg, J.B., Zhang, E.J., Rosenblatt, S. et al. Laserannealing Josephson junctions for yielding scaledup superconducting quantum processors. npj Quantum Inf 7, 129 (2021). https://doi.org/10.1038/s41534021004645
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41534021004645
This article is cited by

Scalable error mitigation for noisy quantum circuits produces competitive expectation values
Nature Physics (2023)

Noisy intermediatescale quantum computers
Frontiers of Physics (2023)

Path toward manufacturable superconducting qubits with relaxation times exceeding 0.1 ms
npj Quantum Information (2022)

Transmon platform for quantum computing challenged by chaotic fluctuations
Nature Communications (2022)

Integrating quantum processor device and control optimization in a gradientbased framework
npj Quantum Information (2022)