Burst-Distance Layout: Placing Error-Correcting Codes Against Anisotropic Phonon Bursts

Superconducting quantum chips sit on silicon or sapphire crystals. When that crystal is under mechanical stress, it occasionally releases a burst of high-energy vibrations (phonons). These phonons break the paired electrons that make the circuit superconducting, and the broken pairs disturb many nearby qubits at once. Error-correcting codes are designed for errors that strike qubits one at a time. A burst instead hits a whole patch of the chip. Whether that patch causes a logical failure depends on which code qubits happen to sit inside it, and that is decided by the compiler that places code qubits onto physical chip positions. Today's placement tools optimize gate count and depth and treat the chip as an abstract graph.

We propose to make placement responsible for burst tolerance. A burst does not spread as a circle. Crystal elastic anisotropy focuses phonons along preferred crystal axes. Our placement tool would therefore assign code qubits so that any single axis-aligned burst footprint covers a set of qubits the code can still correct. Two analogies guide the design. The first is classical interleaving, where bits are spread so that a burst lands on unrelated parity checks. The second is the stress and phonon-transport physics that sets where bursts originate and what shape they take.

Core Claim

The new quantity is a burst distance $d_b$ for a placement $\pi$, which maps code qubits to chip sites. We define it as the minimum number of burst footprints whose combined coverage contains the support of some nontrivial logical operator:

$$d_b(\pi) = \min \{\, m : \exists\, F_1,\dots,F_m \in \mathcal{F},\ \exists\, L \ne I \text{ logical},\ \mathrm{supp}(L) \subseteq \pi^{-1}(F_1\cup\dots\cup F_m) \,\}$$

Here $\mathcal{F}$ is the family of physically allowed footprints. In our model these are ellipses aligned with the substrate's phonon-focusing axes, translated to every chip location, and weighted by a stress map of where bursts originate. $\mathrm{supp}(L)$ is the set of qubits on which logical operator $L$ acts nontrivially.

The ordinary code distance counts single qubits. Burst distance counts physical events. Two placements with identical gate cost can differ sharply in $d_b$. Our claim is that a placement tool that maximizes $d_b$ subject to a routing-overhead budget will lower the logical error rate under realistic burst noise. It should also beat placement that only uses per-qubit error rates, because per-qubit rates cannot express that qubits along one crystal axis fail together.

The physical motivation comes from A stress-induced source of phonon bursts and quasiparticle poisoning. That paper finds that a varnish-glued silicon crystal shows phonon bursts at more than 100 times the rate of a suspended one. The excess decays over 6–10 days. A residual floor remains even in the low-stress mount, which the authors attribute to stress in deposited metal films. The excess quasiparticle density it simulates, about $5\times10^{-8}$, makes these bursts a meaningful error source.

The anisotropic footprint is borrowed from two phonon-transport results. Qian Zhang et al. show that elastic anisotropy focuses heat-carrying phonons along crystal axes. Abhijeet Paul et al. show that phonon conductance in silicon depends on crystal orientation. Neither paper studies qubits. We use them only to justify parameterizing footprint shape by crystal orientation. We treat the actual aspect ratio on a qubit substrate as a quantity to be fitted or swept, not assumed.

Research Program

We would build three components.

The first is a burst noise model. It contains Poisson burst events whose locations are drawn from a stress map concentrated near glue and mount boundaries and under metal films. Each event has an elliptical footprint with orientation set by the substrate cut and an aspect ratio $\alpha$, where $\alpha = 1$ is isotropic. Each event raises the error probability of every qubit inside the footprint for a recovery window. This is layered on top of standard circuit-level noise (the SI-1000 model used by Louvre).

The second is an exact burst-distance evaluator. For moderate codes it searches over footprint combinations against logical operator supports. For larger codes we use a sampled lower bound.

The third is a placement tool that adds a burst-distance constraint to an existing exact formulation. We take the two-stage MaxSAT layout of QECC-Synth, which places data qubits and uses idle "bridge" qubits instead of SWAPs for long-range checks. We add clauses forbidding any logical support from fitting inside $\lceil d^\ast/2\rceil$ footprints, where $d^\ast$ is the target burst distance. For codes too large for MaxSAT, we embed the same penalty in the coarsening and refinement scheme of ML-QLS. The method's role there is to supply scalable placement. The tension is that ML-QLS deliberately clusters interacting qubits, which is exactly the opposite of interleaving. Measuring that tradeoff is a central output.

Code translation enters as the choice of workloads and through one specific question. We compile bivariate bicycle codes, which are classical circulant codes translated into commuting quantum codes, and surface codes. We also compile the entanglement-assisted quasi-cyclic codes of Hsieh, Brun and Devetak. Those codes keep classical girth-6 graphs by using 18 pre-shared entangled pairs on 128 qubits. The entanglement-assisted theory of Brun, Devetak and Hsieh assumes the receiver's halves of these entangled pairs stay error-free. On a chip they are qubits like any other, so the placement tool must decide where they sit. Our objective places them at the lowest-burst-exposure sites. The simulation reports how much logical performance is lost when that assumption is relaxed under burst noise. This is a secondary result, not the core claim.

Baselines

The naive reference is the hand-designed or default geometric layout for each code, with routing as in Louvre for bivariate bicycle codes and the standard lattice for surface codes. This is what a lab would run today.

The strongest existing competitors are QECC-Synth and Louvre. QECC-Synth reports cutting extra CNOT gates by 74.9% on average versus Sabre and by 26.5% versus SATmap. Louvre reports that its 7-step schedule matches the logical error rate of standard degree-6 bivariate bicycle extraction under SI-1000 noise at average connectivity 4.5, while its 8-step schedule costs about 3 times higher logical error at degree 4. Neither reports any result under correlated burst noise. We found no reported result on Lacuna for logical error rate of any layout under phonon-burst noise, so we run both ourselves.

For fairness, every arm uses the same codes, topologies (square grid, heavy-hex, Louvre's degree-4.5 graph), noise samples, and decoder. Every arm also gets the same solver time budget, so the burst-aware MaxSAT receives no extra compute beyond what QECC-Synth is given.

Two ablations isolate the key ingredient. The first is noise-aware placement that gives each qubit an independent error rate equal to its marginal burst exposure under the same stress map. It knows where bursts are likely but not that they are correlated. The second is burst-distance placement with an isotropic footprint ($\alpha = 1$), which isolates the anisotropy.

Beating the baselines means two things at matched or bounded overhead (at most 10% more extra CNOTs than QECC-Synth). The first is a strictly larger $d_b$. The second is a logical error rate per round that is lower by a statistically clear margin under burst noise, with no measurable loss under burst-free SI-1000 noise.

Why not just…?

Why not just use per-qubit noise-aware mapping? Assigning each site an error weight from the stress map is the obvious cheap fix, and it will help by moving qubits away from hot zones. It cannot see that two qubits share a burst. A placement can have low average exposure and still put a weight-$d$ logical operator inside one footprint. This baseline is run directly as an ablation. If it matches burst-distance placement, our correlated objective adds nothing and we would say so.

Why not just mount the chip in a low-stress state? This is a real threat. Anthony-Petersen et al. show suspension cuts the burst rate by more than two orders of magnitude, and good packaging should be done regardless. Their own residual floor from film stress means bursts do not vanish, and the excess decays over days rather than disappearing. We therefore sweep burst rate across the full 100-fold range. If the gain from burst-aware placement disappears at the low-stress rate, the result is that packaging alone suffices, and we would report it.

Why not just let the decoder handle bursts as flagged erasures? If bursts can be detected, a decoder can treat affected qubits as erased. That recovers much of the loss but still fails whenever erased qubits cover a logical support, which is exactly what $d_b$ counts. The two approaches are complementary. We run an arm where the decoder is told burst locations, and test whether burst-aware placement still helps on top of it.

Why not just add anisotropic thermal spreaders to steer phonons away? Heat-spreading layers such as the graphene used in 3D CMOS stacks (Modeling Of 3D Integrated Circuits) and ultra-low cross-plane conductors such as Bi₄O₄SeCl₂ (Gibson et al.) have only been shown at room temperature. Their microwave loss at millikelvin is untested. Placement is a zero-hardware intervention that can be deployed now. If such layers are later shown to work, their effect can be represented in our model as a change in footprint shape.

Proposed Experiments

The main experiment compiles the [[72,12,6]] and [[108,8,10]] bivariate bicycle codes, distance-5 to 9 surface codes, and the n=128 entanglement-assisted codes. Each is compiled onto square-grid and heavy-hex topologies under all baseline and ablation arms. We sweep burst rate (100-fold), footprint size, and aspect ratio $\alpha \in \{1, 2, 4\}$. For each setting we report $d_b$, extra CNOTs, depth, and logical error rate per round.

The decisive checks are three. First, $d_b$ must predict logical error rate across placements with equal gate cost. We expect the error rate to scale roughly as burst rate raised to the power $\lceil d_b/2\rceil$, by the same counting argument that links ordinary distance to error suppression. Second, the anisotropic objective must beat the isotropic ablation when $\alpha > 1$ and tie it when $\alpha = 1$. Third, the gain over the per-qubit noise-aware baseline must survive at the low-stress burst rate and in the arm with burst-flagging decoding.

A failure on the second check would reduce the contribution to correlated-aware placement without the phonon-focusing piece. A failure on the third would show that hardware or decoder fixes make placement unnecessary.

If a multi-sensor cryogenic setup like that of Anthony-Petersen et al. is available, measuring the angle dependence of burst coincidences between sensors would directly calibrate $\alpha$. Without that measurement, $\alpha$ remains a swept parameter.