WorksPhysics and Astronomy
Physics-guided machine learning for sim-to-real calibration of NV diamond magnetometers
Abstract
Ensemble nitrogen-vacancy (NV) centers in diamond enable robust vector magnetometry in unshielded environments, yet deployment remains bottlenecked by complex calibration and a reliance on external data references. Conventional statistical machine learning requires an exorbitantly large volume of training data and suffers from severe simulation-to-reality mismatches. To address this, we introduce a physics-guided hybrid machine learning framework that embeds the Zeeman splitting directly into the learning pipeline. Our physics-guided model significantly reduces the average tracking error demonstrating a 372-fold precision improvement over purely statistical baselines. Furthermore, our hybrid architecture pairs a sparse physical measurement with scalable synthetic data generation, seamlessly incorporating real-world hardware non-idealities. When deployed to decode uncalibrated, raw experimental ODMR data, our framework delivers exceptional predictive accuracy for the scalar magnetic field. This work paves the way toward self-calibrated sensors while establishing a machine learning training method applicable to other data-scarce physical systems
Ensemble nitrogen-vacancy (NV) centers in diamond have matured into a leading solid-state quantum sensing platform due to their atomic-scale spatial resolution[1, 2, 3], wide dynamic range[4], and sub-nT sensitivity[4, 5], some reaching down to [6] under ambient conditions[7, 8, 9]. These properties make them highly appealing for field-ready quantum magnetometers[10, 11, 12, 13] deployed on mobile platforms, defense assets, and remote geophysical survey tools[14]. However, implementing these systems out of controlled laboratory environments introduces severe operational challenges. Signal fluctuations in ambient temperature, local strains, microwave power drifts, and unpredictable background noise degrade the quality of ODMR data acquisition[15, 16, 17]. To maintain metrological accuracy on certain applications, e.g., Position-Navigation-Time (PNT), modern instruments typically rely on periodic manual recalibration or continuous connectivity to external references like GPS or local networks[14]. For NV magnetometers to be field-deployed in unshielded or GPS-denied field environments, sensors must instead possess an internal mechanism to self-calibrate directly from raw, unconditioned experimental data.[18]
Conventional calibration methods use non-linear least-squares fitting algorithms e.g., Levenberg-Marquardt to extract spin-resonance frequencies from ODMR spectra[19, 20]. However, these techniques scale poorly when processing low-contrast, asymmetric[17], or shot-noise-dominated signals[21], and they fail completely without precise initial parameters[22]. Alternative strategies, such as multi-dimensional calibration look-up tables[23] or real-time numerical solvers, bypass these challenges but introduce severe resource bottlenecks, requiring either massive physical storage capacity and/or intensive on-the-fly computation overhead.
Artificial neural networks (ANNs) offer a compelling alternative due to their rapid, millisecond execution times and global convergence capabilities[18, 19, 20]. Using machine learning on NV magnetometer measurement system has shown a speedup of more than 1 order of magnitude with a sequential Bayesian experiment design[24]. In addition, once trained, a neural network acts as a continuous functional mapping that condenses vast physical parameter trajectories into a static, highly optimized weight matrix. This exceptional data compression[25] allows complex quantum behavior to be localized entirely within small memory spaces, making neural architectures uniquely compatible with edge-native microprocessors for field deployment.
Despite these computational advantages, standard deep learning models face limitations when transitioning to physical deployment[26]. Conventionally, these models utilize purely statistical optimization, rendering them highly inefficient, structurally blind to underlying physical laws, and prone to catastrophic failure when operating outside of the training domain and noisy experimental data[27]. This fragility underscores the “simulation-to-reality” (sim-to-real) gap inherent to neural network deployment[28]. While training models exclusively on clean, synthetic data enables rapid optimization, it fundamentally biases the neural network toward idealities that do not exist in practice[19]. Conversely, training a network entirely on experimental measurements is practically unviable as real-world datasets are inherently imperfect and sparse due to the environmental constraints involved in acquiring high-precision, -density dataset. Relying solely on these limited, sparse physical trials leads to severe underfitting, leaving the algorithm blind when confronted with changing field dynamics[27, 29]. Recent efforts have demonstrated the viability of machine-learning-assisted magnetometry[27, 30]. However, these works employed purely statistical trainings without a physical model, which rely on massive, high-density datasets which can cause overfitting.
Here, we introduce a physics-guided machine learning framework optimized for ensemble NV vector magnetometers. By directly embedding the deterministic Zeeman splitting relationship into the training pipeline, our model enforces rapid algorithmic convergence and significantly enhances predictive accuracy. Furthermore, our hybrid approach circumvents physical data scarcity by combining scalable synthetic data generation with sparse empirical measurements, while simultaneously integrating real-world hardware non-idealities into the training pipeline.
The nitrogen-vacancy (NV) center is a point defect formed by a substitutional nitrogen atom adjacent to a lattice vacancy (Figure 1a), whose symmetry axis aligns along one of four tetrahedrally symmetric crystallographic directions , , , and [31]. Because the spin quantization axis of each NV family is fixed along its respective NV bond direction, each family projects a different component of an external magnetic field B onto its axis. The negatively charged defect (NV) forms a spin triplet () in the ground state and it is the magnetically active state relevant for ODMR. The axial projection sensed by the NV family is extracted directly from the ODMR frequency separation, :
Figure 1. NV Center four-axis crystallography, energy diagram and ODMR experimental setup. (a) Atomic structure of the NV center in the diamond lattice, showing four crystallographic NV orientations. (b) Energy diagram of NV-center (c) ODMR Experimental setup.
where GHz/T or 2.8 MHz/G is the electron gyromagnetic ratio. Since the four NV axes are mutually non-coplanar and span three-dimensional space, simultaneous acquisition of ODMR spectra from all four families enables full reconstruction of the three-dimensional magnetic field vector , without mechanical reorientation of the sample[32, 33].
The electron spin-spin interactions within the ground state define the spin quantization axis along the NV symmetry axis, resulting in a zero-field splitting of 2.87 GHz between the and sublevels. These properties allow optically detected magnetic resonance (ODMR), where a microwave field drives transitions between and , producing a characteristic dip in fluorescence intensity at the spin resonance frequency[31]. At zero field, states are degenerate, yielding a single resonance at 2.87 GHz whereas in the presence of an external magnetic field, this degeneracy is lifted via Zeeman splitting of the sublevels into two frequencies whose separation is proportional to the field projected onto the NV symmetry axis (Figure 1b).
As a result, the negatively charged NV center (NV) emits photoluminescence with a zero-phonon line at 637 nm. When optically excited from the spin states, the NV center preferentially relaxes through metastable singlet states before returning to the ground state. In contrast, excitation from the state predominantly undergoes radiative decay through the triplet states. This difference in relaxation pathways enables optical spin initialization into the state and spin-state-dependent fluorescence readout, where reduced photoluminescence intensity indicates population in the states.
The experimental setup (Figure 1c) is based on a confocal optical path designed for ODMR-based NV magnetometry. A 532 nm continuous-wave laser is used for excitation and optical spin initialization of the NV centers in the diamond sample. The laser beam is directed along a free-space beam path and reflected by a dichroic mirror (532 DM) toward the diamond sample. The emitted NV photoluminescence (PL) is collected in reflection geometry and coupled into fiber via a fiber collimator. The collected PL is guided to the photon counting avalanche photodiodes (APD, Excelitas SPCM-NIR), where photon arrival events are registered. Photon timing is recorded by a time-correlated single-photon counting (TCSPC) module (PicoQuant PicoHarp 300), enabling time-resolved PL measurements accessible through the acquisition software. Microwave excitation is delivered to the diamond sample via a wire antenna directly contacted to the sample surface in a pre-wired configuration. A microwave signal generated by a vector signal generator (Rohde & Schwarz SMBV 100A) is applied near 2.87 GHz to drive spin transitions between the and sublevels, producing the characteristic ODMR dip in the PL intensity.
To lift the degeneracy of the four NV orientation families and resolve ensemble Zeeman split resonances, a static magnetic field is applied using a custom-designed three-axis Helmholtz coil assembly, powered by a programmable DC power supply (Rohde & Schwarz HMP4040). A controlled DC current supplied to each coil axis generates a uniform magnetic field at the sample position, inducing Zeeman splitting of the sublevels proportional to the field projection along each NV symmetry axis. The resulting ODMR spectrum, comprising two resonances per NV family, is used to extract the axial field projections and reconstruct the full three-dimensional magnetic field vector[33].
We first generated a simulated dataset by solving the spin Hamiltonian in eq. (2) to extract spin resonances with fixed initial parameters such as the polar and azimuthal angles, yielding clean, symmetric, multi-peak ODMR profiles. We evaluated three optimization strategies implementing a shallow deep learning model utilizing K-fold cross validation (Figure 2). The baseline methodology employs a purely statistical approach, treating the ODMR spectra as high-dimensional input arrays. The architecture utilizes two hidden layers (512 and 128 units), employing Rectified Linear Unit (ReLU) activation functions and a 30% dropout rate and K-fold cross validation to mitigate overfitting. This strategy relies solely on minimizing the empirical loss between predicted and observed spectral features, without adherence to the governing physical constraints. In contrast, our physics-guided framework embeds the end-to-end frequency separation of the ODMR spectral envelope , directly into the learning architecture. By utilizing the peak-finding methods to isolate the resonance lineshapes and fitting them to the deterministic Zeeman splitting relation, , each spectrum is assigned a physics-guided label. This strategy successfully restricts the search space of the optimization manifold, effectively acting as a regularization[34, 35]. Moreover, our ensemble neural network consists of feeding the outputs of the two models into an additional machine learning layer to yield the final predictions. In the next section, we compare the results of these optimization strategies in detail.

Figure 2. Structural Architecture and workflow. The pipeline takes a simulated cw ODMR spectrum as an input array and processes it alongside the true longitudinal magnetic field (). The dataset undergoes a K-fold cross-validation branching into stochastic and physics-guided model. The outputs of the two models are fed into another layer of stochastic model forming an ensemble neural network, yielding the final outputs.
A quantitative comparison of performance metrics, and error distributions among the stochastic model, the physics-guided model, and the ensemble architecture is illustrated in Figure 3. The scatter plots in the first column compare the actual-versus-predicted outputs across 1,000 independent validation samples, visually exposing the progressive optimization gap from the statistical learning to our physicsguided and ensemble models (Figure 3a, Figure 3b, and Figure 3c). For the stochastic model (Figure 3a), the predicted coordinates manifest significant deviations from the true values, illustrating the network's fundamental inability to resolve overlapping spectral lines purely through statistical mapping. In contrast, both the physics-guided model and the ensemble model show clear overlaps between the predicted data points and the actual values across the 0 to 5 G range. As demonstrated by the Mean Squared Error (MSE) convergence profiles across a 5-fold cross-validation (Figure 3d, Figure 3e, and Figure 3f), the stochastic model exhibits severe volatility and an inability to settle into a unified global minimum across folds. Conversely, both the physics-guided and ensemble models achieve immediate, synchronized stability within a few epochs. By directly embedding the end-to-end frequency separation of the ODMR envelope into the learning pipeline, the physics-guided architecture introduces a deterministic constraint governed by the electronic gyromagnetic ratio, suppressing spurious local minima, forcing the optimization path onto stable, physically consistent minimum without adding backpropagation complexity.

Figure 3. Quantitative comparison of performance and average error distributions across the three evaluated network architectures: stochastic (top row), physics-guided (middle row), and ensemble model (bottom row). (a) - (c) Actual-versus-predicted output scatter plots mapped over 1,000 independent evaluation samples. (d) - (f) Mean squared error (MSE) optimization profiles tracking validation loss across a 5-fold cross-validation split over 50 training epochs. (g), (h), (i) Average error and standard deviations plotted as a function of the magnetic field, paired with localized error frequency histograms.
This architectural evolution manifests directly in the average errors (), , and localized error distribution histograms shown in the third column (Figure 3g, Figure 3h, and Figure 3i). The stochastic model (Figure 3g) relies entirely on unconstrained Stochastic Gradient Descent (SGD) to map arbitrary numerical patterns from the raw input distribution. Crucially, in the low magnetic field regime ( G), the stochastic model experiences catastrophic tracking failure, exhibiting massive, unstable residual swings.
By comparison, the physics-guided framework (Figure 3h) dramatically compresses the average error down to G, a two-orders of magnitude improvement over the stochastic model while tightening the standard deviation to G and smoothing out the low-field divergence. Finally, the ensemble model (Figure 3i) shows an improved refinement compressing the average error to and restricts the histogram spread to . Compared to the stochastic model, the physics-guided framework yields a 99.25% reduction in the average error ().
Moving from this physics-guided baseline to the ensemble model yields an additional 64.08% reduction in error, proving that the multi-channel architecture effectively screens out remaining statistical fluctuations. As a result, the ensemble model yields a 99.73% cumulative improvement in precision over the baseline stochastic approach with an improvement factor of 372 (Table 1). While increasing the input data density could theoretically improve interpolation for a purely statistical approach, it would also exacerbate training inefficiencies and increase the computational overhead. Furthermore, in data-sparse regimes, such a strategy typically fails to improve generalization, as the model remains susceptible to capturing non-physical correlations within the spectral noise floor[36].
| Model | Average Error (G or nT) | Standard Deviation (G or nT) | Improvement factor |
| Stochastic | or | or | 1 |
| Physics-Guided | or | or | 133 |
| Ensemble | or | or | 372 |
Table 1. Performance summary for the three evaluated neural network architectures.
However, directly deploying our synthetically optimized neural network onto experimentally measured ODMR data initially failed to recognize the spectra for prediction. Because the synthetic training data lacks real-world experimental non-idealities, such as Gaussian shot noise, power broadening, optical transition asymmetries, and etc., the model faces a severe simulation-to-reality (sim-to-real) domain mismatch when processing uncalibrated measurement data, which is a known vulnerability in unconstrained neural network applications[27, 30]. While mapping a high-density, fully labeled experimental dataset across every physical parameter is practically unfeasible, our hybrid approach circumvents this limitation by utilizing sparse physical measurements as empirical seeds to initialize large-scale data generation capable of capturing both the real-world non-idealities with scalable data generation backed by the underlying Spin Hamiltonian. This approach also prevents the network from developing a simulation bias toward idealized dataset but forces the neural network to adapt directly to the non-idealities inherent in field dynamics.
Our hybrid training pipeline initiates with raw data ingestion, channeling 36 raw experimental ODMR spectra into a parameterized simulation engine using the physical parameters summarized in Supplementary I. After locking in these fitted baseline parameters, the engine scales data generation by sweeping the scalar magnetic field , producing a dense matrix of 14,456 spectra that populate between the experimental data. This expanded matrix is subsequently routed through the neural network training including the K-fold cross validation and hyper-optimization to prevent overfitting, and to determine the best hyper-parameters. The architecture and process flow of our hybrid training pipeline are mapped in Supplement II.
To evaluate performance, the trained network was evaluated using raw ODMR spectra with the measured , that were completely excluded from the training set. Figure 4(a) evaluates the model on uncalibrated, raw experimental ODMR spectra. Predictions cluster tightly along the parity line (y=x), achieving a mean absolute error (MAE) of 7.97 G and root mean square error (RMSE) of 9.52 G. The near-zero systematic bias (, ) confirms that our physics-guided model successfully prevents model drift and overfitting to synthetic baseline artifacts. Furthermore, the individual residual errors tracked in Figure 4(b) remain predominantly within , confirming that the observed spread reflects stochastic variance bounded by intrinsic hardware fluctuations and experimental measurement uncertainties of the uncalibrated testbed rather than systematic model failure.

Figure 4. Experimental validation of the hybrid model evaluated across 20 raw experimental ODMR spectra. (a) Parity plot comparing predicted vs. measured longitudinal magnetic field (). The dashed line indicates an ideal agreement (b) Individual residual prediction errors plotted by sample number.
In summary, we have demonstrated a hybrid, physics-guided machine learning for ensemble nitrogen-vacancy (NV) diamond magnetometers. In the synthetic validation phase, embedding the Zeeman equation directly into the training network allows the physics-guided and the ensemble models to achieve immediate global convergence achieving a 372-fold precision improvement over the stochastic baseline. Furthermore, to bridge the sim-to-real gap, we developed a hybrid training framework that integrates real-world non-idealities with scalable synthetic data generation. By channeling a sparse, noisy experimental spectra into a parameterized simulation engine, the pipeline locked baseline parameters and expanded the dataset across a 0 to 120 G dynamic range. These results demonstrate that our hybrid framework successfully overcomes the interpolation limit of synthetic-only training.
Ultimately, our physics-guided, hybrid neural network architecture possesses a broad versatility that extends beyond nitrogen-vacancy diamond magnetometry, offering a generalizable framework for diverse physical measurement systems where acquiring high-density, high-precision data remains a critical bottleneck. The governing physical laws and constraints of an arbitrary system can be structurally implemented into the physics-guided framework while the parameterized hybrid data generation can successfully bypass the scarcity of physical training data. This dual-method establishes a scalable paradigm for edge-deployable neural networks optimized for a wide spectrum of metrological applications.
Acknowledgement
This work was supported by the National Science Foundation (NSF) under the Centers of Research Excellence in Science and Technology (CREST) program (Grant No. HRD-1914777) and the IITP (Institute of Information & Communications Technology Planning & Evaluation) grant (RS-2025-25464832)
Conflict of Interest
The authors have no conflicts to disclose.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
References
- [1]S. Hong, M.S. Grinolds, L.M. Pham, D. Le Sage, L. Luan, R.L. Walsworth, and A. Yacoby, “Nanoscale magnetometry with NV centers in diamond,” MRS Bull. 38(2), 155–161 (2013).DOI
- [2]L. Rondin, J.-P. Tetienne, P. Spinicelli, C. Dal Savio, K. Karrai, G. Dantelle, A. Thiaville, S. Rohart, J.-F. Roch, and V. Jacques, “Nanoscale magnetic field mapping with a single spin scanning probe magnetometer,” Appl. Phys. Lett. 100(15), 153118 (2012).DOI
- [3]T.X. Zhou, R.J. Stöhr, and A. Yacoby, “Scanning diamond NV center probes compatible with conventional AFM technology,” Appl. Phys. Lett. 111(16), 163106 (2017).DOI
- [4]H. Clevenson, L.M. Pham, C. Teale, K. Johnson, D. Englund, and D. Braje, “Robust high-dynamic-range vector magnetometry with nitrogen-vacancy centers in diamond,” Appl. Phys. Lett. 112(25), 252406 (2018).DOI
- [5]F.M. Stürner, A. Brenneis, T. Buck, J. Kassel, R. Rölver, T. Fuchs, A. Savitsky, D. Suter, J. Grimmel, S. Hengesbach, M. Förtsch, K. Nakamura, H. Sumiya, S. Onoda, J. Isoya, and F. Jelezko, “Integrated and Portable Magnetometer Based on Nitrogen-Vacancy Ensembles in Diamond,” Adv. Quantum Technol. 4(4), 2000111 (2021).
- [6]Y. Xie, H. Yu, Y. Zhu, X. Qin, X. Rong, C.-K. Duan, and J. Du, “A hybrid magnetometer towards femtotesla sensitivity under ambient conditions,” Sci. Bull. (Beijing). 66(2), 127–132 (2021).DOI
- [7]G. Balasubramanian, I.Y. Chan, R. Kolesov, M. Al-Hmoud, J. Tisler, C. Shin, C. Kim, A. Wojcik, P.R. Hemmer, A. Krueger, T. Hanke, A. Leitenstorfer, R. Bratschitsch, F. Jelezko, and J. Wrachtrup, “Nanoscale imaging magnetometry with diamond spins under ambient conditions,” Nature 455(7213), 648–651 (2008).DOI
- [8]J.R. Maze, P.L. Stanwix, J.S. Hodges, S. Hong, J.M. Taylor, P. Cappellaro, L. Jiang, M.V.G. Dutt, E. Togan, A.S. Zibrov, A. Yacoby, R.L. Walsworth, and M.D. Lukin, “Nanoscale magnetic sensing with an individual electronic spin in diamond,” Nature 455(7213), 644–647 (2008).DOI
- [9]J.F. Barry, J.M. Schloss, E. Bauch, M.J. Turner, C.A. Hart, L.M. Pham, and R.L. Walsworth, “Sensitivity optimization for NV-diamond magnetometry,” Rev. Mod. Phys. 92(1), 15004 (2020).DOI
- [10]B. Zhao, H. Guo, R. Zhao, F. Du, Z. Li, L. Wang, D. Wu, Y. Chen, J. Tang, and J. Liu, “High-Sensitivity Three-Axis Vector Magnetometry Using Electron Spin Ensembles in Single-Crystal Diamond,” IEEE Magn. Lett. 10, 1–4 (2019).DOI
- [11]J.M. Schloss, J.F. Barry, M.J. Turner, and R.L. Walsworth, “Simultaneous Broadband Vector Magnetometry Using Solid-State Spins,” Phys. Rev. Appl. 10(3), 34044 (2018).
- [12]B. Chen, X. Hou, F. Ge, X. Zhang, Y. Ji, H. Li, P. Qian, Y. Wang, N. Xu, and J. Du, “Calibration-Free Vector Magnetometry Using Nitrogen-Vacancy Center in Diamond Integrated with Optical Vortex Beam,” Nano Lett. 20(11), 8267–8272 (2020).DOI
- [13]M. Tsukamoto, K. Ogawa, H. Ozawa, T. Iwasaki, M. Hatano, K. Sasaki, and K. Kobayashi, “Vector magnetometry using perfectly aligned nitrogen-vacancy center ensemble in diamond,” Appl. Phys. Lett. 118(26), 264002 (2021).DOI
- [14]Ryan S. Cassel, William G. Tobias, and Bonnie L. Marlow, “PR-23-0577-Quantum-vs-Classical-Complementary-PNT,” MITRE, (2023).
- [15]M. Simanovskaia, K. Jensen, A. Jarmola, K. Aulenbacher, N. Manson, and D. Budker, “Sidebands in optically detected magnetic resonance signals of nitrogen vacancy centers in diamond,” Phys. Rev. B 87(22), 224106 (2013).DOI
- [16]K. Hayashi, Y. Matsuzaki, T. Taniguchi, T. Shimo-Oka, I. Nakamura, S. Onoda, T. Ohshima, H. Morishita, M. Fujiwara, S. Saito, and N. Mizuochi, “Optimization of Temperature Sensitivity Using the Optically Detected Magnetic-Resonance Spectrum of a Nitrogen-Vacancy Center Ensemble,” Phys. Rev. Appl. 10(3), 34009 (2018).
- [17]A. Jarmola, Z. Bodrog, P. Kehayias, M. Markham, J. Hall, D.J. Twitchen, V.M. Acosta, A. Gali, and D. Budker, “Optically detected magnetic resonances of nitrogen-vacancy ensembles in 13C-enriched diamond,” Phys. Rev. B 94(9), 94108 (2016).DOI
- [18]M. Krenn, J. Landgraf, T. Foesel, and F. Marquardt, “Artificial intelligence and machine learning for quantum technologies,” Phys. Rev. A (Coll. Park). 107(1), 10101 (2023).DOI
- [19]V. Cimini, I. Gianani, N. Spagnolo, F. Leccese, F. Sciarrino, and M. Barbieri, “Calibration of Quantum Sensors by Neural Networks,” Phys. Rev. Lett. 123(23), 230502 (2019).DOI
- [20]V. Dunjko, and H.J. Briegel, “Machine learning & artificial intelligence in the quantum domain: a review of recent progress,” Reports on Progress in Physics 81(7), 074001 (2018).arxiv.org/abs/1709.02779
- [21]K. Hayashi, Y. Matsuzaki, T. Taniguchi, T. Shimo-Oka, I. Nakamura, S. Onoda, T. Ohshima, H. Morishita, M. Fujiwara, S. Saito, and N. Mizuochi, “Optimization of Temperature Sensitivity Using the Optically Detected Magnetic-Resonance Spectrum of a Nitrogen-Vacancy Center Ensemble,” Phys. Rev. Appl. 10(3), 34009 (2018).
- [22]M. Fujiwara, A. Dohms, K. Suto, Y. Nishimura, K. Oshimi, Y. Teki, K. Cai, O. Benson, and Y. Shikano, “Real-time estimation of the optically detected magnetic resonance shift in diamond quantum thermometry toward biological applications,” Phys. Rev. Res. 2(4), 43415 (2020).DOI
- [23]B.N. Wenny, and K. Thome, “Look-up table approach for uncertainty determination for operational vicarious calibration of Earth imaging sensors,” Appl. Opt. 61(6), 1357–1368 (2022).DOI
- [24]S. Dushenko, K. Ambal, and R.D. McMichael, “Sequential Bayesian Experiment Design for Optically Detected Magnetic Resonance of Nitrogen-Vacancy Centers,” Phys. Rev. Appl. 14(5), 54036 (2020).DOI
- [25]G.C. Maríno, A. Petrini, D. Malchiodi, and M. Frasca, “Deep neural networks compression: A comparative survey and choice recommendations,” Neurocomputing 520, 152–170 (2023).DOI
- [26]J.P. Cohen, T. Cao, J.D. Viviano, C.W. Huang, M. Fralick, M. Ghassemi, M. Mamdani, R. Greiner, and Y. Bengio, “Problems in the deployment of machine-learned models in health care,” CMAJ 193(35), E1391–E1394 (2021).DOI
- [27]27 J. Homrighausen, L. Horsthemke, J. Pogorzelski, S. Trinschek, P. Glösekötter, and M. Gregor, “Edge-Machine-Learning-Assisted Robust Magnetometer Based on Randomly Oriented NV-Ensembles in Diamond,” Sensors 23(3), (2023).DOI
- [28]28 D.L. Craig, H. Moon, F. Fedele, D.T. Lennon, B. van Straaten, F. Vigneau, L.C. Camenzind, D.M. Zumbühl, G.A.D. Briggs, M.A. Osborne, D. Sejdinovic, and N. Ares, “Bridging the Reality Gap in Quantum Devices with Physics-Aware Machine Learning,” Phys. Rev. X 14(1), 11001 (2024).DOI
- [29]29 M.C. Caro, H.-Y. Huang, N. Ezzell, J. Gibbs, A.T. Sornborger, L. Cincio, P.J. Coles, and Z. Holmes, “Out-of-distribution generalization for learning quantum dynamics,” Nat. Commun. 14(1), 3751 (2023).
- [30]30 M. Tsukamoto, S. Ito, K. Ogawa, Y. Ashida, K. Sasaki, and K. Kobayashi, “Accurate magnetic field imaging using nanodiamond quantum sensors enhanced by machine learning,” Sci. Rep. 12(1), 13942 (2022).DOI
- [31]31 L. Rondin, J.-P. Tetienne, T. Hingant, J.-F. Roch, P. Maletinsky, and V. Jacques, “Magnetometry with nitrogen-vacancy defects in diamond,” Reports on Progress in Physics 77(5), 056503 (2014).
- [32]32 L.M. Pham, D. Le Sage, P.L. Stanwix, T.K. Yeung, D. Glenn, A. Trifonov, P. Cappellaro, P.R. Hemmer, M.D. Lukin, H. Park, A. Yacoby, and R.L. Walsworth, “Magnetic field imaging with nitrogen-vacancy ensembles,” New J. Phys. 13, (2011).DOI
- [33]33 S. Steinert, F. Dolde, P. Neumann, A. Aird, B. Naydenov, G. Balasubramanian, F. Jelezko, and J. Wrachtrup, “High sensitivity magnetic imaging using an array of spins in diamond,” Review of Scientific Instruments 81(4), (2010).DOI
- [34]34 J.J. Pannell, S.E. Rigby, and G. Panoutsos, “Physics-informed regularisation procedure in neural networks: An application in blast protection engineering,” International Journal of Protective Structures 13(3), 555–578 (2022).DOI
- [35]35 M.A. Nabian, and H. Meidani, “Physics-Driven Regularization of Deep Neural Networks for Enhanced Engineering Design and Analysis,” J. Comput. Inf. Sci. Eng. 20(1), (2019).DOI
- [36]36 K. Donhauser, A. Tifrea, M. Aerni, R. Heckel, and F. Yang, Interpolation Can Hurt Robust Generalization Even When There Is No Noise (Adv. Neural Inf. Process. Syst., 2021).arxiv.org/abs/2108.02883
