NVRNet: Deep Learning Model for Fast Nitrogen Vacancy Characterization under Room Temperature

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 $0.44$-$0.67\times$ 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.

Figures (14)

• Figure 1: Workflow of NVRNet. A Ramsey PL trace is collected under room temprature with single photon detector. Then it is first processed by the denoising module, which comprises a simulation-pretrained core augmented with uncertainty-aware adapter layers, fine-tuned on experimental traces. The resulting denoised trace is then passed to the parameter-estimation network, which predicts the ${}^{13}\mathrm{C}$ count (up to $n_{\max}=9$) and the associated parallel hyperfine couplings $A_{\parallel}$.
• Figure 2: Comparison between experimental and simulated noise statistics and principal components. (a) Distributions in PL(%) for experimental (blue) and synthetic (pink) residuals. (b) PCA embedding of full traces: the simulated dataset (pink) spans the region occupied by the experimental data (blue) in the leading two principal components, indicating that simulation covers the dominant experimental variability.
• Figure 3: Core of the denoising network. Stage 1 denoises the Fourier-domain representation and returns to the time domain. Stage 2 refines in the time domain using a UNet with a self-attention bottleneck to preserve long-range phase coherence and weak beating.
• Figure 4: Adapters for experimental fine-tuning. A front adapter ingests the noisy PL trace and an uncertainty channel and produces the core input. The core, composed of a pretrained UNet, is frozen during fine-tuning. A back adapter maps the output to the experimental trace format.
• Figure 5: Experimental denoising via simulation pretraining and adapter fine-tuning.(a--c) Example Ramsey PL(%) traces from three NV centers. Black: raw experimental input. Orange: 200-sweep fitted reference ($+10$ PL(%) offset). Green: simulation-trained core-only output ($-10$ PL(%) offset). Blue: core+adapter output after experimental fine-tuning. Offsets are for visualization only. (d) RMSE distributions (in PL percentage points) over all held-out few sweep test traces ($K\le 10$) for each NV, comparing the raw experimental baseline (orange), the simulation-trained core-only model (green), and the experimentally fine-tuned core+adapter model (blue). (e) RMSE versus number of averaged sweeps $K$ per inference.