Fidelity-informed neural pulse compilation of a continuous family of quantum gates with uncertainty-margin analysis
Arash Fath Lipaei, Ebrahim Khaleghian, Selin Aslan, Gani Göral, Zidong Lin, Özgür E. Müstecaplıoğlu
Abstract
We develop a fidelity-informed neural pulse-compilation framework for a continuous family of single-qubit gates on a three-qubit liquid-state nuclear magnetic resonance (NMR) processor. Instead of decomposing each target unitary into a sequence of calibrated basis gates, the method learns a direct map from the axis-angle parameters of an arbitrary U_2 in SU(2) operation to a piecewise-constant radio-frequency control sequence that implements the desired transformation. Training is performed end-to-end through the time-ordered propagator of the driven Hamiltonian using global-phase-insensitive unitary fidelity as the learning signal. We show numerically that a single model generalizes across a continuous range of gate parameters and experimentally validate representative compiled pulses on a benchtop three-qubit NMR device. In addition, we analyze sensitivity to structured perturbations in Hamiltonian and control parameters by introducing a prescribed uncertainty set and performing a comparative risk-aware redesign based on right-tail Conditional Value-at-Risk (RU-CVaR). This stage produces pulse solutions with broader tolerance margins within the chosen uncertainty model. The results demonstrate continuous pulse-level gate synthesis in an experimentally accessible setting and illustrate a hardware-aware compilation strategy that can be extended to other quantum platforms. While the uncertainty model considered here is tailored to NMR, the neural compilation and risk-aware optimization framework are general and may be useful in architectures where calibration overhead, parameter drift, or control constraints make repeated per-gate optimization costly.
AI Impact Assessments
(3 models)Scientific Impact Assessment
Core Contribution
This paper presents a neural network-based pulse compiler that maps continuous gate parameters (axis-angle representation of SU(2)) directly to piecewise-constant RF control sequences for a three-qubit liquid-state NMR processor. The key idea is replacing discrete gate decomposition with a single neural model that generalizes across a continuous family of single-qubit gates. The secondary contribution is a risk-aware re-optimization stage using Right-tail Conditional Value-at-Risk (RU-CVaR) to broaden tolerance margins against structured Hamiltonian and control perturbations.
The approach builds on prior work by Sauvage & Mintert (2022) on optimal control of gate families and Khaleghian et al. (2026) on GRAPE-based neural pulse generation for NMR, extending these ideas with the continuous compilation framework and the CVaR-based robustness layer.
Methodological Rigor
Strengths in formulation: The physics-informed training pipeline is well-constructed. The end-to-end differentiable propagator, global-phase-insensitive fidelity metric, and trigonometric feature encoding are all appropriate choices. The gradient computation through the time-ordered propagator via Fréchet derivatives is standard but correctly implemented.
Weaknesses in experimental validation: The experimental validation is limited. Only a single gate instance (θ=90°, γ=90°, α=10°) receives full state tomography, yielding 92% fidelity—a notable gap from the >99% numerical predictions. The authors acknowledge this gap is due to state-preparation errors, finite coherence times, RF inhomogeneity, and model mismatch, but do not quantify these contributions. The spectral-phase linearity check (Figure 7) provides supporting evidence but is a relatively weak diagnostic compared to full process tomography across the gate family.
Numerical study limitations: The numerical evaluation is restricted to rotations within [0, 90°]³, which covers only a fraction of SU(2). The authors explicitly acknowledge this is a "controlled proof-of-principle region," but this significantly limits the demonstrated generality. The training set consists of only 512 gates, and validation on a 3° mesh within this restricted domain, while showing good coverage, does not stress-test the approach on the full rotation group.
Uncertainty analysis: The RU-CVaR framework is borrowed from financial risk management and applied in a straightforward manner. The prescribed uncertainty model is reasonable for NMR but, as the authors repeatedly emphasize, is not experimentally calibrated. The robustness claims are therefore conditional on the assumed perturbation model. The comparison between different α values (0.2, 0.5, 0.8) in Table 6 is informative, showing the expected trade-off between nominal performance and worst-case tolerance.
Potential Impact
The practical impact for NMR quantum computing is modest. Liquid-state NMR quantum processors are not competitive platforms for scalable quantum computation, and the benchtop device used (SpinQ Triangulum Mini) is primarily an educational/research tool. The paper's repeated framing that the approach "may be useful" for superconducting, Rydberg, or other platforms remains speculative without any demonstration or concrete analysis on those systems.
The conceptual contribution—neural compilation of continuous gate families with risk-aware optimization—has some value but is incremental over existing literature. Neural pulse optimization and differentiable quantum control have been explored extensively. The CVaR-based robustness layer adds a useful perspective from stochastic optimization but is applied in a relatively straightforward manner without novel theoretical insights about the control landscape structure.
The most useful aspect for the broader community may be the structured uncertainty analysis framework, which provides a template for how to systematically probe sensitivity of learned pulses to different error channels.
Timeliness & Relevance
The paper addresses a genuine need: variational quantum algorithms and parameterized circuits increasingly require continuously parameterized gates, and per-gate calibration is costly. However, the field has already seen significant work on neural and ML-based quantum control (including the authors' own prior work). The NMR setting limits the timeliness, as the community's attention has shifted toward superconducting and neutral-atom platforms where these challenges are more acute.
Strengths & Limitations
Key strengths:
Notable weaknesses:
Additional Observations
The paper would benefit significantly from: (1) a direct comparison with GRAPE-optimized pulses for the same gate family, (2) extension to the full SU(2) manifold, (3) multi-qubit gate compilation, and (4) transfer experiments to a different platform. The writing is careful but sometimes overly defensive, spending significant text managing expectations rather than deepening the technical analysis.
Generated Apr 19, 2026
Comparison History (50)
Paper 1 presents a highly practical, hardware-validated machine learning approach to quantum pulse compilation with robust uncertainty analysis. Its potential to reduce calibration overhead and improve gate fidelity makes it broadly applicable and impactful across various quantum hardware platforms. In contrast, Paper 2 relies on classical simulations and focuses on a specific theoretical condensed matter model, making its immediate real-world impact narrower.
Paper 1 targets a core bottleneck for scalable fault-tolerant quantum computing—real-time QEC decoding under strict latency constraints—and provides architecture unification plus a deployability-focused FPGA compression/quantization pipeline with concrete microsecond-scale guidance. This is timely and broadly relevant across quantum hardware efforts pursuing surface codes, with clearer near-term systems impact. Paper 2 is innovative (continuous gate-family pulse synthesis with risk-aware robustness) and experimentally validated, but demonstrated on a niche NMR platform and primarily affects compilation/control workflows rather than the central fault-tolerance stack, likely yielding narrower cross-field impact.
Paper 2 likely has higher impact: it tackles fault tolerance at the system-architecture level for distributed quantum computing, addressing a central barrier to scalable quantum machines—robust operation under node/device failure and hot-swapping. The results connect directly to quantum error correction, networking, and hardware modularity, with broad relevance across platforms and near-term timeliness as modular quantum networks emerge. Paper 1 is novel and experimentally grounded, but is demonstrated on single-qubit gates in NMR and may have narrower cross-field reach; ML pulse compilation is active yet often platform-specific despite generalizable ideas.
Paper 1 likely has higher impact due to a more actionable, platform-relevant contribution: a learned continuous gate-to-pulse compiler with experimental validation and an added risk-aware robustness analysis (RU-CVaR) addressing calibration drift/uncertainty. This directly targets a major bottleneck in near-term quantum hardware control and can transfer to other platforms, giving strong real-world applicability and timeliness. Paper 2 offers elegant theoretical insights linking complementarity to geometric/thermodynamic response in open systems, but its immediate methodological and translational leverage appears narrower without clear experimental demonstrations or tooling.
Paper 1 addresses the fundamental challenge of entanglement spectrum estimation for large quantum many-body systems with a novel hybrid quantum-classical framework that combines variational SVD with classical orthogonality correction. It tackles multiple key problems (barren plateaus, hardware noise, measurement complexity) simultaneously and demonstrates broad applicability across 1D and 2D systems. Paper 2, while technically sound, addresses a narrower problem (pulse compilation for single-qubit gates on NMR) with more limited scope and platform specificity. Paper 1's impact spans quantum computing, condensed matter physics, and tensor network methods, giving it broader reach.
Paper 1 makes a fundamental theoretical contribution by extending quantum state tomography with minimal cumulative disturbance to arbitrary finite-dimensional pure states, resolving whether low-regret tomography is a qubit-specific phenomenon or a general geometric property. The O(d³ log² T) regret bound is a strong result with broad implications for quantum information theory. Paper 2, while practically useful, addresses a more incremental engineering problem—neural pulse compilation for NMR—with limited generalizability beyond its specific platform. Paper 1's mathematical depth, generality, and conceptual novelty give it significantly higher potential for broad scientific impact.
Paper 2 offers higher potential impact by combining theoretical innovation in machine learning with experimental validation on actual quantum hardware. Addressing pulse-level compilation and hardware uncertainty tackles a major bottleneck in scaling quantum computers, providing a generalizable framework with broader practical applicability across different architectures compared to the purely numerical, algorithm-focused approach in Paper 1.
Paper 2 presents a more novel and broadly applicable framework—neural pulse compilation with risk-aware optimization (CVaR)—that addresses a practical problem (calibration overhead, parameter drift) across multiple quantum hardware platforms. It combines machine learning, optimal control theory, and uncertainty quantification in a unified framework with experimental validation. Paper 1, while methodologically sound, reports largely negative/expected results (simple Z-basis measurements suffice, classification fails above ~12 qubits) with limited actionable insights. Paper 2's techniques have clearer paths to real-world adoption in quantum computing hardware development.
Paper 2 integrates machine learning with quantum control, offering a highly practical, hardware-aware framework for pulse compilation that reduces calibration overhead. Its cross-disciplinary approach, experimental validation, and generalizability to various quantum platforms suggest a broader and more immediate real-world impact compared to Paper 1's purely theoretical mathematical proof, despite the latter's rigorous foundational value.
Paper 2 is more novel and broadly applicable: it introduces an end-to-end neural pulse compiler that generalizes over a continuous family of gates, validated experimentally, and adds a risk-aware (RU-CVaR) uncertainty-margin redesign—highly relevant to scalable, drift-tolerant quantum control and compilation across platforms. Its methodological rigor (propagator-based training, experiments, and robustness analysis) and real-world utility (reducing calibration/per-gate optimization overhead) suggest wider cross-field impact (ML, control, hardware compilation). Paper 1 is timely for NISQ optimization but is largely simulation-based and closer to incremental improvements over existing variational/annealing heuristics.
Paper 2 provides a rigorous mathematical proof for the error scaling of universally robust dynamical decoupling sequences (URn), filling a significant theoretical gap for a widely used technique in quantum control. This foundational result has broad impact across all quantum computing platforms that use dynamical decoupling, establishing the construction on firm analytical grounds. Paper 1 presents a useful but incremental engineering contribution—neural pulse compilation for NMR with uncertainty analysis—that is more platform-specific and methodologically combines existing techniques (neural networks, CVaR optimization) rather than establishing fundamental new results.
Paper 1 offers a more novel, practically actionable contribution: a learned, pulse-level compiler that generalizes over a continuous SU(2) gate family, experimentally validated on hardware, and extended with risk-aware (CVaR) robustness to uncertainties—highly relevant to reducing calibration/optimization overhead in near-term quantum control. Its methodology is end-to-end physics-informed and includes uncertainty-margin analysis, supporting rigor and real-world utility across platforms. Paper 2 provides useful negative/limitation results and theory for circuit-family classification with measurement strategies, but impact is narrower (small-qubit ML classification) and less directly enabling for scalable hardware performance.
Paper 1 presents a novel neural pulse compilation framework with risk-aware optimization (RU-CVaR) for quantum gate synthesis, combining machine learning with quantum control in a methodologically rigorous way. It addresses a fundamental challenge in quantum computing—efficient, hardware-aware gate compilation—with experimental validation on real hardware. Paper 2 proposes a relatively straightforward preprocessing heuristic for TSP that, while useful, represents an incremental contribution to a well-studied problem. Paper 1's cross-platform generalizability, novel uncertainty analysis, and relevance to the rapidly growing quantum computing field give it substantially higher impact potential.
Paper 2 is more conceptually novel and broadly impactful: it proposes a new architectural primitive (“quantum actuators”) that could reshape how globally controlled quantum processors achieve selectivity, connectivity, and long-range entanglement without local control. Its applicability spans multiple hardware platforms and links to quantum thermodynamics (quantum batteries), increasing cross-field reach and timeliness for scalable architectures. Paper 1 is rigorous and experimentally validated, but is narrower (single-qubit SU(2) pulses on liquid-state NMR) and more incremental within quantum optimal control/compilation; its real-world impact may be limited by platform relevance.
Paper 1 has higher potential impact due to its practical, hardware-demonstrated methodology for continuous, pulse-level quantum gate compilation with a learning-based approach and explicit robustness (RU-CVaR) to model uncertainties—directly relevant to scalable quantum control and compilation across platforms. It offers clear real-world applications (reducing calibration/optimization overhead, handling drift/constraints) and bridges ML, control, and experimental quantum hardware. Paper 2 provides valuable theoretical/numerical insights into FV scaling in open quantum spin chains, but its applications are more specialized and less immediately enabling for quantum technologies.
Paper 2 presents a more novel and broadly applicable framework combining neural pulse compilation with risk-aware optimization (RU-CVaR) for quantum gate synthesis. It addresses fundamental challenges in quantum computing—calibration overhead, parameter drift, and control constraints—relevant across multiple quantum platforms. The continuous gate family compilation and uncertainty-margin analysis represent methodological innovations with wider impact potential. Paper 1, while practical, applies a relatively standard block coordinate descent technique to a specific application (portfolio optimization) on niche hardware, offering more incremental contributions to a narrower community.
Paper 2 likely has higher impact due to stronger real-world applicability and timeliness: it demonstrates experimentally validated, hardware-aware neural pulse compilation for a continuous family of gates, addressing practical calibration/overhead issues in quantum control. The addition of uncertainty-margin analysis (RU-CVaR) improves robustness, a key need for near-term devices, and the framework is plausibly transferable across platforms. Paper 1 is novel and useful for generative QML data encoding, but its scope is narrower (1D distributions, encoding effects) and is less directly tied to immediate experimental quantum advantage.
Paper 1 presents a significant methodological advance for computing NMR properties of paramagnetic molecules by unifying existing formalisms and providing a computationally efficient route that bypasses complex g-tensor and ZFS calculations. This addresses a longstanding challenge in paramagnetic NMR spectroscopy with broad applications in chemistry and materials science. Paper 2, while technically sound, addresses a more incremental improvement in quantum gate compilation for NMR quantum computing—a platform with limited scalability. Paper 1's impact spans computational chemistry, spectroscopy, and materials characterization, offering broader and more immediate real-world utility.
Paper 2 addresses a fundamental open problem in quantum information theory—computing one-way distillable entanglement beyond known special cases. It introduces novel mathematical conditions (regularized less-noisy, informationally degradable), proves new additivity results, and proposes a generalized spin-alignment principle with broad theoretical implications. These contributions advance core quantum information theory and connect to quantum channel capacity questions, impacting a wide community. Paper 1, while technically solid, presents an incremental engineering contribution (neural pulse compilation for NMR) with narrower scope and limited novelty beyond combining known techniques (neural networks, CVaR optimization) in a specific experimental setting.
Paper 2 proposes a novel quantum simulation scheme for the Motzkin spin chain using Rydberg atoms, connecting fundamental concepts in condensed matter physics (topological phases, area-law violations) and high-energy physics (AdS/CFT). It addresses a computationally hard problem with a concrete experimental proposal on a rapidly advancing platform. Paper 1, while technically solid, is more incremental—applying neural network compilation and risk-aware optimization to NMR pulse design, a relatively mature and limited-scalability platform. Paper 2 has broader cross-disciplinary impact and higher novelty in bridging mathematical physics with experimental quantum simulation.