Model-Uncertainty-Aware Digital Twins for Non-Markovian Quantum Repeater Memories

Abstract

Quantum-repeater controllers need reliable estimates of stored entanglement as adjacent links are generated. Fixed exponential lifetimes miss non-Markovian revivals, while complex surrogates are unnecessary when the microscopic family is known. We present QMemTwin, a physics-constrained digital twin that accounts for model uncertainty. It calibrates a bank of contractive passive state-space models, combines their amplitudes with finite-sample-corrected Akaike weights, and abstains when temporal-holdout data indicate unsupported dynamics. Passivity guarantees a completely positive, trace-preserving amplitude-damping channel for each candidate and their convex amplitude average. We evaluate [N cases] synthetic memories: [N supported] supported passive time-invariant cases and [N structural] nonstationary structural-shift cases. In the matched one-pseudomode regime, the correctly specified one-pseudomode fit performs best. Across [N unknown] supported cases with uncertain families, QMemTwin reduces population RMSE by [population reduction]% and Bell-fidelity RMSE by [Bell reduction]% relative to the two-pseudomode model; both paired effects remain significant after Holm correction. The flexible passive model has a slightly lower mean error for unknown families but is statistically indistinguishable from QMemTwin and loses accuracy in the matched regime. The temporal-holdout gate detects [detection]% of structural shifts, with [false alarm]% false alarms on supported families and AUROC [AUC]. In 2-, 4-, and 8-link repeater chains, the mixed cohort shows no significant 8-link goodput gain. A predeclared high-backflow stress cohort yields an 8-link goodput gain of [goodput gain]% over the two-pseudomode model, with paired p = [p value]. These results favor using a known family when available, accounting for model uncertainty otherwise and abstaining on unsupported dynamics.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.