[论文解读] NVRNet: Deep Learning Model for Fast Nitrogen Vacancy Characterization under Room Temperature

Chao Shang, Gregory D. Fuchs|arXiv (Cornell University)|Mar 14, 2026

Diamond and Carbon-based Materials Research被引用 0

一句话总结

NVRNet 提供一个物理信息驱动的仿真到现实世界的管道，去噪噪声 Ramsey 律迹并直接估计 NV 中心在室温下的 13C 超精细耦合， enabling 快速 NV 表征。（保留原文含义）

ABSTRACT

Characterization of the local spin environment of single diamond nitrogen-vacancy centers is a critical task for quantum sensing, quantum networking, and diamond materials optimization. We introduce NVRNet, a physics-informed simulation-to-reality pipeline that maps a fast acquisition, noisy Ramsey photoluminescence (PL) trace to a denoised waveform as well as outputting a direct estimate of hyperfine coupling to ${}^{13}\mathrm{C}$ spins in the environment. The denoiser is a two-stage time-frequency UNet followed by an attention-augmented time-domain UNet, pretrained on Hamiltonian-based simulations with experimentally calibrated noise. The simulation-pretrained, experimentally fine-tuned denoiser reduces the median reconstruction error on held-out few-sweep experimental traces to $0.44$-$0.67 imes$ that of the raw experimental noisy traces across the three NV centers. A transformer-based estimator trained on simulation labels then predicts hyperfine parameters, and forward reconstruction from the inferred parameters reproduces the dominant experimental time- and frequency-domain features, with representative normalized FFT reconstruction errors of 0.10-0.19. These results establish NVRNet as a fast, hardware-compatible route to hyperfine inference from minimal Ramsey data.

研究动机与目标

• Motivate rapid, robust characterization of NV centers and their local nuclear spin environments under realistic lab noise.
• Develop a simulation-to-reality pipeline that leverages synthetic data to train a denoising network and a hyperfine-parameter estimator.
• Integrate a two-stage denoiser with uncertainty adapters to handle experimental drift and noise not captured by simulations.
• Produce physically interpretable hyperfine parameters (13C count and parallel couplings) from minimal Ramsey data.

提出的方法

• Two-stage denoising: (i) frequency-domain coarse denoising with a 1D CNN–UNet, (ii) time-domain refinement with an attention-augmented UNet.
• Self-attention is added at UNet bottlenecks to capture long-range phase coherence in Ramsey traces.
• Denoiser is pretrained on Hamiltonian-based simulations with calibrated noise and finetuned on experimental traces via lightweight uncertainty adapters.
• Hyperfine parameter estimator is transformer-based and trained on simulation labels to predict 13C count (0–10) and parallel couplings, with forward reconstruction validating inferred parameters.
• Oscillation-pattern encoders extract global waveform factors (T2*, overall amplitude, and detuning) fed into a Transformer head for hyperfine prediction.
• Noise model for synthetic data includes shot noise, drift, and PL fluctuations, tuned to match experimental residual statistics.

实验结果

研究问题

• RQ1 Can a physics-informed ML pipeline denoise few-sweep Ramsey traces better than raw data under realistic lab noise?
• RQ2 Can a simulation-trained denoiser, augmented with lightweight experimental adapters, generalize to real NV traces?
• RQ3 Can a Transformer-based estimator accurately infer 13C count and parallel hyperfine couplings from denoised Ramsey traces?
• RQ4 Does forward reconstruction using inferred hyperfine parameters reproduce key time- and frequency-domain features of the experiment?

主要发现

NV	n	Base med.	Core med.	Core+Adapt med.
NV 1	40000	2.255	1.570	0.984
NV 2	40000	1.864	3.392	1.248
NV 3	40000	1.897	2.021	1.856

• Adapter-finetuned denoiser reduces median reconstruction error to 0.44–0.67× that of raw traces on held-out NV centers.
• Denoiser improves denoising performance in the few-sweep regime (K ≤ 10) across multiple NV centers.
• Hyperfine estimator predicts 13C count and parallel couplings with representative forward-reconstruction fidelity, enabling reliable initialization for further refinement.
• Forward reconstruction using inferred hyperfine parameters reproduces dominant time- and frequency-domain features with FFT reconstruction errors of 0.10–0.19.
• Simulation-pretrained core with experimental adapters outperforms simulation-only models in handling real-world noise.