Correlated radiation erasures limit cross-chip erasure codes in modular superconducting processors: a simulation study
Cross-chip erasure codes [Xu et al., PRL 129, 240502 (2022)] encode each chip's logical qubit in an inner surface code and combine chips into an outer erasure code. A radiation burst that erases one chip is then a located, correctable error, and the scheme's failure-rate formula assumes that chips fail independently. We test that assumption by simulation. Geant4 transport through a model refrigerator feeds a deposit-to-loss chain calibrated on Google's gap-engineered Willow processor, using a burst-to-error mapping that Google measured with an echo pulse in the stabilizer circuit. The chain does not transfer to a second gap-engineered array we tested, so it applies to Willow-class chips only. In a measured two-chip device, a burst on one chip coincided with one on the other 2.4% of the time (8 events). With one detector threshold fitted, and one Geant4 setting corrected after a first comparison, our transport predicts 1.5%. We consider a [[4,1,2]] outer code of four data chips and an ancilla chip, with distance-7 inner codes and a 1 ms outer correction cycle, on chips mounted flat 1 to 3 mm apart in one sample box. Counting line gammas and muons at a measured laboratory gamma flux, these layouts fall 140 to 370 times short of the formula's protection, the upper end from nine-chip arrays. They fall 26 to 60 times short at the higher flux that Willow's burst rate implies. In our model, single events that erase several chips, mostly through material the chips share, cause the gap. Placing each of the five chips in its own sample box at least 5 cm apart cuts the block's correlated failure rate 40 to 60 times at inner-code distance 7 (95% intervals down to 24 times). The comparison is with the same chips on one carrier, 1 mm apart. With lost chips located perfectly, the separated block then recovers 15 to 25% of the formula's protection on our interpolated loss curve. Located from burst detections by the simplest attribution rules, under our nominal loss model, it recovers 0.3 to 0.5%. On chips using the echo, writing off every chip with a burst detected at the qubits' own threshold does worse than no outer code when all five chips share a carrier. These detection results assume that bursts too small to detect do not destroy the inner logical qubit, which our simulation cannot resolve. All rates isolate radiation, the floor that remains once intrinsic logical errors are suppressed.
Authors
- Henderson Dylan (ORCID: https://orcid.org/0009-0009-7022-4986)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24412996
- Primary Topic
- Quantum Computing Algorithms and Architecture
- Type
- preprint