🔬 applied physics

Machine-learning based high-bandwidth magnetic sensing

This paper addresses the sensitivity-bandwidth tradeoff in nitrogen-vacancy (NV) center magnetic sensing by implementing machine learning tools that reduce the required data points by at least a factor of three while maintaining current error levels, thereby advancing the feasibility and efficiency of quantum sensing technologies.

Original authors: Galya Haim, Stefano Martina, John Howell, Nir Bar-Gill, Filippo Caruso

Published 2026-10-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Galya Haim, Stefano Martina, John Howell, Nir Bar-Gill, Filippo Caruso

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the microscopic world of quantum physics, scientists have long sought ways to use the strange behavior of atoms to measure the invisible forces that shape our reality. One of the most promising tools for this task is a tiny defect found inside diamonds, known as a nitrogen-vacancy center. Imagine a place in the diamond's crystal structure where a carbon atom is missing and has been replaced by a nitrogen atom sitting next to an empty spot. This specific arrangement acts like a sensitive antenna for magnetic fields. When researchers shine a green laser on these defects and bathe them in microwaves, the defects glow with a red light. The brightness of this glow changes depending on the strength and direction of the magnetic field around it. By carefully tuning the microwaves and watching how the light changes, scientists can map out magnetic fields with incredible precision. This technique has opened doors to studying everything from ancient rocks to the activity of living cells. However, there is a significant hurdle: to get a complete picture of a magnetic field, especially one that varies wildly in strength, the traditional method requires scanning through a vast range of frequencies, collecting a massive amount of data point by point. This process is slow, and the more data needed, the longer the measurement takes, which limits how fast scientists can track changing magnetic signals.

A team of researchers has now found a way to speed up this process dramatically without losing the precision needed for accurate measurements. Instead of relying on the slow, step-by-step scanning method that has been the standard for years, they turned to machine learning, a branch of artificial intelligence that allows computers to learn patterns from data. The researchers trained a computer program, specifically a type of neural network, to recognize the unique signatures of magnetic fields from a much smaller set of information. In their experiments, they used both real measurements taken from diamond samples and computer-generated simulations of what those measurements should look like. They taught the computer to predict the exact location of the magnetic field's resonance points by looking at only a fraction of the data points that a traditional scan would require.

The results were striking. The machine learning model was able to reconstruct the full magnetic field map using as few as one-tenth of the data points that the traditional method usually demands. Even with this drastic reduction in data, the error in the measurement remained low, staying below one megahertz. In fact, the researchers found that they could cut the measurement time by a factor of three while keeping the same level of accuracy. This is a crucial improvement because it means scientists can now track magnetic fields that change rapidly over time, something that was previously difficult or impossible with the slower scanning techniques. The study showed that the computer model was not just faster; it was also more robust. When the data was noisy or the magnetic signals were overlapping and hard to distinguish, the traditional scanning method often failed to find all the necessary points. The machine learning model, however, could still identify the correct patterns, effectively seeing through the confusion where the old method stumbled.

To ensure their findings were reliable, the team tested their approach under various conditions, changing the noise levels and the width of the signals in their simulations. They discovered that for the computer model to work well in the real world, it needed to be trained on a diverse set of examples, including data with different levels of noise and different signal shapes. When the model was trained on a wide variety of scenarios, it became flexible enough to handle the messy, unpredictable nature of real-world experiments. The researchers also compared their method against the traditional approach using actual data collected from their laboratory equipment. While the machine learning model performed slightly worse than the traditional method when using a full set of data points, it vastly outperformed the traditional method when the data was sparse. This suggests that the best approach might be a hybrid one, where the computer is trained to handle specific amounts of data, allowing researchers to choose the right tool for the job depending on how much information they can gather.

The implications of this work extend beyond just making measurements faster. By breaking the trade-off between sensitivity and speed, this technique could revolutionize how we study dynamic systems. Whether it is monitoring the electrical activity in a circuit, tracking transient signals in biological tissues, or characterizing materials that change over time, the ability to gather high-quality magnetic data in a fraction of the time opens new possibilities for scientific discovery. The researchers demonstrated that by combining the physical precision of quantum sensors with the pattern-recognition power of machine learning, we can overcome the limitations that have held back quantum sensing for years. This is not a magic trick or a sudden leap into the future, but a practical engineering solution that makes existing technology more efficient and capable. The work confirms that with the right training, a computer can learn to see the magnetic world more clearly and quickly than the traditional tools ever could, paving the way for more advanced and responsive quantum technologies.

Technical Summary: Machine-learning based high-bandwidth magnetic sensing

Problem Statement
Nitrogen-vacancy (NV) centers in diamond are a leading platform for high-sensitivity, high-spatial-resolution magnetic sensing. However, standard spin resonance measurement techniques, specifically Optically Detected Magnetic Resonance (ODMR) via Electron Spin Resonance (ESR), face a fundamental trade-off between sensitivity, dynamic range, and bandwidth. In large dynamic range scenarios, identifying the vectorial magnetic field requires scanning the entire relevant frequency space to locate eight distinct resonance frequencies (corresponding to the four crystallographic orientations of the NV center). This "raster scanning" approach necessitates a high number of data points to resolve resonance positions, particularly when resonances are narrow or the Signal-to-Noise Ratio (SNR) is low. Consequently, measurement time increases, limiting the ability to track time-varying signals and reducing overall sensitivity (defined as $\eta = \delta B \sqrt{T}$ ). The authors identify the need for a method to reduce the required number of data points without compromising measurement error or sensitivity.

Methodology
The authors propose a Quantum Machine Learning (QML) approach using a classical feed-forward neural network, specifically a Multilayer Perceptron (MLP), to reconstruct the external magnetic field vector directly from sparse ESR data.

Model Architecture: The network is designed as a regression problem. The input layer consists of $n$ measured data points ( $\lambda_i$ ) from an ESR spectrum, and the output layer predicts the eight resonance frequencies ( $y_i$ ). The network utilizes Rectified Linear Unit (ReLU) activation functions for hidden layers and a sigmoid function for the output layer.
Training Strategy: The models were trained using a combination of real experimental data and simulated data generated from a simplified NV Hamiltonian. The simulation accounted for Zeeman shifts, Lorentzian line shapes, Gaussian noise, and varying resonance widths.
Optimization: Hyperparameters (learning rate, batch size, dropout, weight decay, and network depth/width) were optimized using the Ray Tune platform with Bayesian optimization (Hyperopt) and the ASHA scheduler. The training objective was to minimize the Mean Squared Error (MSE) between predicted and ground-truth resonance frequencies using the Adam optimizer.
Experimental Setup: Real data was acquired using an epi-illumination wide-field microscope setup with a 532 nm green laser and microwave pulses. The dataset included 96 full raster scans under varying magnetic fields.
Comparison Baseline: The ML performance was benchmarked against standard raster scanning with Lorentzian curve fitting. Crucially, the comparison accounted for the statistical probability of successfully identifying all eight resonances in subsampled data. The error for raster scans was normalized by the square root of the success probability ( $\sqrt{P}$ ) to fairly compare it with the ML model, which outputs a prediction regardless of whether resonances are visually distinct.

Key Results

Data Reduction and Error Scaling: The ML models demonstrated the ability to maintain low measurement error with significantly fewer data points compared to raster scanning. Specifically, the network achieved similar error levels with only 200 data points as the raster scan did with 600 points, representing a threefold reduction in measurement time while preserving sensitivity.
Performance at Low Data Counts: With only 10% of the data points (60 out of 600), the ML model maintained an error below 1 MHz. In contrast, raster scanning error increased significantly at similar subsampling rates due to the failure to identify all eight resonances.
Robustness to Overlapping Resonances: The neural network successfully predicted all eight resonance frequencies even in cases where Lorentzian peaks overlapped or were degenerate (e.g., data samples with only 6 distinct peaks). Standard raster scanning often failed in these scenarios, particularly at lower SNRs.
Generalization and Training Strategies:
- Dataset Size: Training on larger datasets (100,000 samples) yielded lower errors than smaller sets (1,000 samples), though 10,000 samples provided a substantial improvement over the baseline.
- Robustness to Variations: Networks trained on data with a range of Lorentzian widths (5–16 MHz) and SNR values (2.5–10) were significantly more robust to variations in test data than networks trained on specific, fixed parameters.
- Interpolation vs. Specific Training: While training a single network on full-length data and interpolating subsampled inputs saves training time (as one network fits all datasets), it comes at the cost of less favorable performance. Conversely, training specific networks for specific subsampling ratios yields lower average errors and standard deviations.
Real Data Validation: When tested on 46 real data samples, the ML models showed higher errors than on simulated data (indicating some overfitting), but still outperformed raster scanning in terms of error scaling and robustness, particularly at lower data point counts.

Significance and Claims
The paper claims that integrating machine learning into NV-based quantum sensing offers a viable path to overcoming the sensitivity-bandwidth trade-off in large dynamic range scenarios. The primary contribution is the demonstration that neural networks can reconstruct vectorial magnetic fields with high fidelity using a fraction of the data points required by traditional raster scanning.

The authors state that this approach enables a potential reduction in required data points by at least a factor of three while maintaining current error levels. This improvement directly translates to enhanced sensitivity for time-varying signals or allows for constant sensitivity with even faster measurement times. The work promotes the use of quantum machine learning protocols to make quantum sensing technologies more feasible and efficient, specifically for applications involving large, dynamic signals such as circuit characterization and transient biological signal detection. The authors emphasize that the success of this method relies on accurate physical simulations for training data and training on diverse datasets to ensure robustness against experimental noise and line-shape variations.

Technical Summary: Machine-learning based high-bandwidth magnetic sensing

More like this