Quantum sensors that compute: quantum computational magnetic-field sensing using a superconducting qubit

AI summary

Overview

Research area: Quantum sensing and quantum computing with superconducting circuits — specifically quantum computational sensing (QCS), which merges a sensing protocol with an in-situ quantum computation using quantum signal processing (QSP).

Technical level: Intermediate. The core idea is explained conceptually, but full engagement with the protocol design, benchmarking, and the mutual-information analysis assumes familiarity with qubits, Bloch-sphere dynamics, Ramsey interferometry, and dynamical decoupling.

Scope: An experimental demonstration, on a single flux-tunable superconducting transmon, that interleaving trained control unitaries with magnetic-field sensing intervals yields higher binary-classification accuracy than optimized conventional sensing baselines given the same total signal access time.

What This Paper Is About

A single measurement of a one-qubit sensor returns at most one bit of information, so the usual route to classifying a sensed signal is to repeat a sensing protocol many times, build a high-precision estimate of the field, and then compute the class label classically. That approach spends sensing time acquiring a full parameter estimate even when only a coarse task-relevant property (such as a class label) is needed. This paper asks whether a single superconducting qubit can instead perform the computation itself — shaping the sensed signal directly into the measurement outcome — and whether that beats carefully optimized conventional baselines under matched sensing-time budgets.

Key Contributions

1. Experimental realization of QCS on one device. Sensing and computing were both performed by the same single flux-tunable superconducting transmon, with the magnetic field sensed through the double-junction SQUID loop that forms part of the qubit. The task-relevant function is encoded into the single bit revealed by each shot, with classical postprocessing reduced in nearly all cases to a single threshold.

2. Protocol extension and noise-aware training. The authors adapted the QSP-based binary-classification protocols theoretically proposed in Ref. [12] to continuously applied signals by accounting for the signal-induced detuning during control operations, and optimized them in the presence of stochastic phase noise. Control unitaries were trained in silico on a digital twin of the system via supervised learning.

3. Benchmarking against optimized conventional baselines across a task-complexity family. Six labeling rules of increasing complexity were used, defined by class boundaries $c_{n,k} = \frac{k}{n+1} + \frac{1}{20}\sin\left(\frac{2\pi k}{n+1}\right)$ and labels $F_n^*(u) = k \bmod 2$. Baselines were a multitime Ramsey phase-estimation protocol with linearly spaced interrogation times (chosen over exponentially spaced times because the latter produce binary subdivisions misaligned with the class boundaries) for static fields, and dynamical-decoupling (DD) sequences for oscillating fields, each optimized over accessible control parameters and shot allocations.

4. A mechanistic account of the advantage. Cosine similarity between measured excitation probability $P_e$ and the target function, together with mutual-information tracking through the QSP sequence, shows how task information present early in the Bloch vector is concentrated onto the measurement axis by the end of the protocol, while nuisance information (e.g. random phase) is suppressed.

Main Findings

• Static fields: 3.0 to 14.9 percentage points advantage. At a fixed total sensing time of 100 µs, QCS maintained a consistent accuracy advantage of 3.0 to 14.9 percentage points over the conventional Ramsey baseline plus classical postprocessing, with larger advantages on more complex tasks. The conventional baseline itself came within 3.0 percentage points of its ideal, shot-noise-limited performance, so the comparison is not against a weakened straw man.

• QCS beat the ideal, shot-noise-limited baseline too. Across all but the simplest task, the experimental QCS protocol outperformed the baseline under ideal, shot-noise-limited conditions for both protocols, with gains from 3.4 to 10.3 percentage points.

• Oscillating fields, amplitude classification: 3.8 to 20.1 percentage points. In the phase-incoherent regime ($\omega/2\pi$ fixed at 2 MHz, amplitude $A$ swept from 0 to 5 MHz, random phase per shot), QCS beat optimized DD across the full 0 to 200 µs range, with improvements of 3.8 to 20.1 percentage points and accuracies from 85.3% to 96.5%. The advantage persisted across readout fidelities from 50% to perfect readout, indicating the gain is not a readout artifact.

• Oscillating fields, frequency classification: up to 15.4 percentage points, but not universally. For signals with a 1 MHz bandwidth centered at 1.5 MHz and $A$ fixed at 0.6 MHz, gains reached 15.4 percentage points, while in the worst case performance decreased by 2.8 percentage points; accuracies ranged from 81.9% to 93.8%. QCS underperformed DD on the simplest (single-threshold) frequency task and won on the more complex tasks with multiple disjoint intervals, consistent with DD being naturally suited to narrow-band discrimination.

• Depth is traded against repetitions. Pareto-optimal protocols shift from shallow, single-shot operation to deeper, repeatedly measured operation as the time budget grows. Static tasks: low-budget $L = 2$–$14$ with $S = 4$–$7$, high-budget $L = 4$–$16$ with $S = 29$–$46$. AC amplitude: $L = 5$–$7$, $S = 1$ rising to $L = 8$–$22$, $S = 42$–$68$. AC frequency: $S = 1$ with $L = 5$–$7$ rising to $L = 12$–$28$, $S = 32$–$43$.

• Expressivity scales with circuit depth. Cosine similarity between measured $P_e$ and the target function increases with the number of interleaved control layers, and the measured $|g\rangle$/$|e\rangle$ distributions sharpen at the class-transition boundary. The number of layers needed for a given error grows approximately linearly with task number up to about 18, beyond which gains become marginal.

• Information is concentrated, not merely acquired. Early in the QSP sequence, mutual information about the task is present in the full Bloch vector but not accessible to a single projective measurement; by the end, most of it is concentrated in the measurement basis, while amplitude information about $u$ and phase information about $\phi$ are largely absent there.

Methodology in Plain English

The sensor is a transmon qubit whose transition frequency depends on magnetic flux through a SQUID loop. The qubit is deliberately biased away from its flux sweet spot and driven on resonance at that bias point, so that any flux-induced frequency shift detunes the drive and causes the qubit to accumulate phase — an effective Z rotation whose angle records the sensed field.

The protocol alternates short sensing intervals (where the signal imprints phase) with programmed control rotations, repeated for $L$ layers. All parameters are fixed in advance from a simulation of the device, trained by supervised learning to maximize the correlation between finitely sampled measurement outcomes and the true class labels. The sequence is then repeated for $S$ shots, the binary outcomes are averaged, and a decision rule — in nearly all cases a single threshold implemented with a random-forest classifier — returns the class. Total sensing time $T_s$ counts everything over all shots: control pulses, free-evolution intervals, and readout, since the signal stays on during readout and readout settings are fixed.

For fairness, baselines were optimized as aggressively as the QCS protocol: Ramsey interrogation times, shot allocations, and DD pulse counts and interpulse delays (up to 500 ns, with the six best sequences selected for experiment) were all tuned per task. Protocols were compared via Pareto fronts (best accuracy achievable at a given time budget), selected on a validation set, and evaluated on an independent test set, with error bars from 20 computational resamplings of the measurement outcomes.

The device was a double-junction flux-tunable transmon (aluminum on silicon) with $\omega_q = 2\pi \times 5.125$ GHz, $\omega_r = 2\pi \times 5.618$ GHz, $T_1 = 30$ µs, $T_2^* = 8$ µs, ground- and excited-state readout errors of 29% and 27%, and flux sensitivity near the bias point of approximately 29 MHz/mA. Measurements were performed in a Bluefors LD250 dilution refrigerator at roughly 9 mK.

Why This Matters

The result shows that a computational advantage in sensing does not require a large quantum processor: a single noisy qubit with 8 µs of dephasing and percent-level readout errors already outperforms optimized conventional protocols when the task — not the parameter — is the target. It reframes the sensing figure of merit from estimation precision to task accuracy, and demonstrates that measurement-induced information loss, not just decoherence, is a resource that can be engineered around.

Real-world applications the authors point to:

• NV-center magnetic sensing for biomedical applications, where task-specific readout could shorten acquisition.
• SQUID-based magnetoencephalography, where classification of field signatures rather than full field reconstruction is often the actual goal.
• Atomic-vapor magnetometers for navigation, where fast, low-latency decisions matter more than high-fidelity field values.
• Spatially distributed sensing, since the multiqubit extension the authors sketch would let a single measurement encode answers about signals at several locations.

Industry relevance spans superconducting quantum hardware providers (the protocol runs on a standard tunable transmon with existing control stacks), quantum-sensing instrument makers, and medical-imaging and navigation-metrology companies that care about sensing time and latency rather than raw precision.

Future Directions

• Tasks beyond binary classification. Multiclass problems motivate multiple qubits, so that a single measurement can carry enough information to encode the answer; multiqubit settings also permit multidimensional inputs (spatially distributed signals) and more expressive computations, including multivariable polynomial transformations. Open questions include efficient multiqubit training and robustness to realistic noise.

• Adaptive protocols. The present study is restricted to non-adaptive protocols because of experimental setup limitations. Comparing QSP-based QCS against conventional baselines in an adaptive setting — where earlier outcomes steer later shots — is flagged as an important next step.

• More realistic task classes. The advantages shown are for synthetic classification tasks; identifying which practical sensing problems admit a QSP-based QCS advantage remains open.

• Porting to other platforms. The authors argue practical QCS will need a platform that is simultaneously a state-of-the-art sensor for the task and coherent and controllable enough to execute the protocol faithfully, naming NV centers, SQUID systems, and atomic-vapor magnetometers as candidate hosts.

Target Audience

Quantum-sensing experimentalists and superconducting-qubit researchers; quantum-information scientists interested in the boundary between metrology and computation; signal-processing and machine-learning researchers exploring learned quantum control; and graduate students or technically fluent readers who want a concrete, benchmarked example of computation folded into a sensing protocol on near-term hardware.

Authors’ abstract

A measurement of a single-qubit quantum sensor reveals at most 1 bit of information about the signal that was sensed. To perform a classification task on the signal, the conventional approach is to repeat a sensing protocol many times, averaging the measurement results to obtain a high-precision estimate of the sensed signal, and then to apply classical postprocessing. Quantum computational sensing (QCS) is an alternative approach that breaks with the paradigm of first obtaining a classical estimate of the signal and then computing a function of the estimated signal in postprocessing. QCS instead combines quantum sensing with quantum computing to concentrate information about the signal relevant to the task into the solitary bit revealed by each measurement. Here, we report on the experimental demonstration of QCS where sensing and computing were both performed by the same single superconducting transmon qubit. We consider various binary classification tasks based on static and oscillating magnetic fields induced by current through a flux line. The fields were sensed through a double-junction superconducting quantum interference device (SQUID) loop that was part of the qubit. We used a protocol based on quantum signal processing to preprocess the sensed signals in the quantum domain prior to measurement. For tasks on static magnetic fields, our protocol outperformed the conventional baseline of Ramsey-based phase estimation by as much as 15 percentage points. For tasks on oscillating magnetic fields, we classified signal amplitude and frequency with up to 20 and 15 percentage points higher accuracy, respectively, compared to the conventional baseline of optimized dynamical-decoupling protocols. Our results illustrate how quantum computing can enhance quantum sensing even with a minimally sized quantum system subject to the practical limitations of decoherence and error-prone operations.