Automating detection of Two-Level Systems in Superconducting Qubits

AI summary

Overview

Research area: Superconducting quantum computing hardware — specifically, materials-level decoherence from microscopic two-level system (TLS) defects in transmon qubits, and automated analysis of time-resolved SWAP spectroscopy data.

Technical level: Intermediate (accessible with basic familiarity with qubit coherence, T₁ relaxation, and spectroscopy; the computer-vision machinery is explained in plain terms).

Scope: The paper introduces and benchmarks two automated pipelines for detecting TLS defects in SWAP spectroscopy maps, then uses them to compare 52 transmon qubits with and without moderate (~10%) frequency trimming via Alternating-Bias Assisted Annealing (ABAA).

What This Paper Is About

Superconducting transmon qubits lose coherence partly because of microscopic TLS defects in amorphous oxides and junction barriers, which show up as avoided crossings ("chevrons") in time-resolved SWAP spectroscopy. Extracting reliable defect statistics from large numbers of qubits by hand, or by reducing each measurement to a one-dimensional decay-rate trace, is a bottleneck. The paper's goal is a deterministic, non-parametric, automated way to localize these defects directly on the native two-dimensional spectroscopy maps, validated against conventional fitting, and then applied to ask whether defect populations can be modified after fabrication.

Key Contributions

1. A deterministic 2D computer-vision pipeline (2d-cv) that localizes TLS defects directly on the full SWAP spectroscopy matrix Z(t, f), using anisotropic Gaussian scale-space background estimation (σ_t = 80, σ_f = 10 pixels), positive-bounded "residual darkness" mapping, a global percentile threshold θ = P₈₅(D), and temporal persistence projection ρ(f) with a threshold ρ_th = 0.14 — with no iterative nonlinear optimization and no supervised training.
2. A rigorous cross-benchmark against conventional 1D decay-rate fitting (1d-drf), using nearest-neighbor spectral matching within δf = 25 MHz, yielding n = 597 consensus defect pairs and Pearson correlation r = 0.88 on log-transformed frequency-integrated loss strengths across approximately six orders of magnitude.
3. A unified physical loss metric ℐ_TLS = πAγ (units Hz²) obtained by fitting a Lorentzian excess-relaxation model to the same 1d-drf-derived spectrum for candidates from either pipeline, making the two methods comparable in physical units; the derivation links ℐ_TLS = (π/2)v⊥² to the square of the transverse qubit–TLS coupling.
4. The first large-scale lateral statistical study of TLS landscapes after post-fabrication annealing, across 52 flux-tunable transmon qubits split into an untrimmed control ensemble (n = 28) and an ABAA-trimmed ensemble (n = 24, ~10% frequency trimming), plus a junction-geometry-stratified robustness check across nine discrete SQUID junction-area designs.

Main Findings

• The two pipelines agree quantitatively. For n = 597 consensus pairs matched within δf = 25 MHz, frequency-integrated loss strengths ℐ_TLS = πAγ agree with Pearson correlation r = 0.88 on log-transformed values, spanning approximately six orders of magnitude with no evident systematic offset from the 1:1 line.
• Total defect count is unchanged by moderate annealing. 1d-drf gives 38.7 ± 17.7 defects/qubit (untrimmed) versus 37.9 ± 11.3 (trimmed); 2d-cv gives 40.8 ± 17.2 versus 40.1 ± 16.2. Two-sample t-tests confirm these shifts are statistically insignificant (p > 0.8).
• Per-qubit counts vary widely. Individual qubits span roughly 10 to 85 defects/qubit, so device-to-device spatial heterogeneity dominates count variance, with a mean of about 39 defects/qubit.
• Total dielectric loss drops by more than a factor of two. Span-integrated loss ℒ = Σ Γ₁^TLS(f_i) Δf falls from 6.9 kHz to 3.1 kHz after moderate (~10%) ABAA trimming. A single outlier at 76 kHz exists in the baseline ensemble; excluding it gives a baseline mean of ≈4.5 kHz and still a statistically significant reduction of over 30%.
• The suppression is selective, not uniform. The loss-strength distribution for untrimmed qubits is continuous and unimodal (log-normal-like) from ~10⁷ to 10¹³ Hz². After trimming, the high-dissipation tail above ℐ_TLS > 10¹¹ Hz² is systematically depleted, shifting the ensemble toward weaker coupling.
• Count–loss decoupling. Total detectable TLS density stays statistically constant while the span-integrated loss halves — interpreted as selective weakening of the most strongly dissipative defect channels rather than removal of defect states.
• The mechanism is coupling suppression, not linewidth broadening. Because ℐ_TLS = (π/2)v⊥², the loss reduction reflects attenuation of the transverse dipole coupling matrix element v⊥ itself, consistent with bias-driven atomic reorganization in the AlₓOᵧ barrier via a Cabrera–Mott mechanism.
• Junction geometry is ruled out as the cause. Loss reduction persists within individual junction-area categories among the nine SQUID designs.
• The methods are complementary, not redundant. 2d-cv-exclusive detections capture weakly dissipative fluctuators and coherent chevrons with T₁^TLS/T₁^bg ≈ 0.85–0.95; consensus detections correspond to severe T₁ bottlenecks.

Methodology in Plain English

Each qubit is prepared in its excited state with a π-pulse, then a flux pulse of amplitude A and duration t tunes its frequency across a spectral window; a dispersive readout records the excited-state population, producing a 2D map of population versus frequency and time.

The 1d-drf pipeline takes each frequency column, fits a single exponential P(t) = A e^(−t/τᵢ), converts it to a relaxation rate Γ₁(fᵢ) = 1/τᵢ, estimates a smooth background with a rolling 15th-percentile filter over a window of W = 50 frequency channels, estimates the noise floor with a rolling median absolute deviation scaled by 1.4826, and flags positive excursions above baseline + 1.5σ as defect candidates.

The 2d-cv pipeline instead works on the whole map. It builds a smooth background by blurring the map with a Gaussian that is wide in time and narrow in frequency, matched to the shape of a chevron (spectrally narrow, temporally long). It then measures "darkness" — how much the actual population falls below that background, with anything above background clipped to zero. It thresholds that darkness globally at the 85th percentile (a global rather than per-column threshold avoids forcing false positives into flat regions), then tallies for each frequency channel the fraction of the timeline that is active, ρ(f). Frequency channels with ρ(f) > 0.14 are merged into defect regions, the defect frequency is assigned to the channel of maximum persistence, and an integrated darkness score S ranks prominence.

To compare the two methods in physical units rather than pipeline-specific statistics, both sets of candidates are refit with the same Lorentzian model on the same rate spectrum, giving ℐ_TLS = πAγ, retaining only fits with amplitude A ≥ σ_local. Candidates from the two methods are matched if they lie within 25 MHz, and matched features are split into consensus, 2d-cv-exclusive, and 1d-drf-exclusive sets.

Why This Matters

Impact on research: The paper provides an automated, open-source, deterministic alternative to hand-tuned or supervised-ML defect characterization, and gives the first large-scale statistical evidence that moderate ABAA treatment selectively suppresses strongly coupled TLS channels while conserving total defect density. It reframes TLS defects as partially engineerable rather than immutably fixed by amorphous disorder.

Real-world applications:

• High-throughput automated characterization and calibration of multi-qubit superconducting processors.
• Wafer-scale quality assurance and fabrication process control for quantum foundries.
• Feedback into annealing/trimming recipes that target the most dissipative defect channels.
• Continuous, automated retuning routines during processor operation.

Industry relevance: The datasets come from Rigetti processors, and the framework is explicitly positioned as a scalable diagnostic asset for hardware vendors and for processor optimization in fault-tolerant architectures. Because the pipeline relies on general spatiotemporal coherence rather than device-specific calibration models, it is claimed to be extensible to broader multi-qubit architectures and wafer-scale protocols.

Future Directions

• Extending the 2d-cv pipeline to other multi-qubit architectures and to continuous automated retuning, given its independence from device-specific calibration.
• Establishing whether stronger or differently tuned ABAA treatments push beyond ~10% trimming to change total defect count, not just the high-dissipation tail.
• Connecting the observed v⊥ attenuation to specific microscopic structural changes in the AlₓOᵧ barrier, which the paper proposes but does not resolve structurally.
• Comparing the deterministic computer-vision approach directly against machine-learning regression on spectroscopic maps, which the paper positions itself against but does not benchmark numerically.
• Evaluating the consequences of the count–loss decoupling for actual coherence times and gate fidelities of the trimmed devices, which the paper does not report.

Target Audience

Experimental quantum-computing groups working on superconducting hardware, materials and device physicists studying decoherence and TLS defects, fabrication and process engineers at quantum hardware companies, and data-analysis researchers interested in non-parametric computer-vision methods for noisy scientific imaging. Readers focused on gate-level benchmarking or algorithmic quantum computing will find the metrics here device-diagnostic rather than performance-oriented.

Authors’ abstract

Microscopic two-level system (TLS) defects remain a primary mechanism of decoherence and operational instability in superconducting transmon qubits, necessitating scalable and automated methods for their characterization. Here, we present and benchmark two complementary analysis pipelines for extracting TLS statistics directly from time-resolved SWAP spectroscopy: one-dimensional decay-rate fitting (1D-DRF), which detects defects via localized enhancements in the qubit relaxation rate, and a deterministic, non-parametric computer-vision framework (2D-CV) that achieves two-dimensional spectral localization by exploiting the temporal persistence of coherent population suppression. We deploy both methods on SWAP spectroscopy measurements from 52 flux-tunable transmon qubits on Rigetti processors with and without moderate ($\sim 10\%$) post-fabrication frequency trimming via Alternating-Bias Assisted Annealing (ABAA). We show that both pipelines converge on a consistent global characterization of the defect landscape while exhibiting complementary sensitivity across distinct coupling regimes. Crucially, both methods independently reveal a count--loss decoupling under moderate annealing: while the total detectable TLS defect density remains statistically unchanged, the span-integrated dielectric loss is reduced by approximately a factor of two, demonstrating selective suppression of the most strongly dissipative defect channels. These results establish an automated, non-parametric analysis framework for high-throughput hardware diagnostics and provide a statistical baseline for post-fabrication defect engineering in large-scale superconducting quantum processors.