The term "signal deconvolution" gets used loosely in instrument engineering. I want to describe specifically what we do in DeteQt's processing pipeline and, equally specifically, where the method does not solve the problem it might appear to solve. Honest signal processing documentation is rare in hardware companies because it exposes limitations. But a potential customer who misunderstands what deconvolution can do will be disappointed in the field, and that is a worse outcome than transparency.
The raw data problem in NV-center magnetometry
An NV-center sensor returns a photoluminescence intensity spectrum as a function of applied microwave frequency. The ODMR spectrum shows dips at the spin resonance frequencies, and the separation between these dips encodes the local magnetic field magnitude. In a perfect instrument with no noise, extracting the field from the spectrum is a peak-finding problem. In a real instrument, three overlapping interference terms complicate the measurement.
First, photon shot noise adds a statistical scatter across the spectrum proportional to the square root of the detected count rate. Second, the NV ensemble in a real diamond chip has a distribution of local strain environments, which broadens the resonance lines and reduces contrast, making precise peak location harder. Third, environmental magnetic fields at frequencies within the measurement bandwidth add to the geological signal in ways that are not distinguishable from the sensor output alone.
The third term is the one deconvolution directly addresses. The first and second terms set the fundamental sensitivity floor that no processing can recover below.
What the deconvolution pipeline does
Our pipeline operates in two stages. The first stage uses the reference gradiometer channel (a second NV sensor physically positioned to have minimal sensitivity to the local geological gradient) to estimate the ambient interference field time series. This estimated interference is subtracted from the primary channel data before spectrum fitting. The subtraction is not perfect, because the two channels have different positions and slightly different spectral sensitivities, but it reduces broadband interference by 25 to 35 dB across the 1 to 50 Hz band in typical field conditions.
The second stage uses a trained sequence model to separate the residual interference from the geological gradient signal. The model was trained on a dataset of 1,840 field and controlled-environment measurements with known geological and interference components, spanning magnetite-bearing gabbro, sulphide-altered basalt, barren greenstone, and several industrial EMI environments. The model learns temporal and spectral features that distinguish geological gradient (spatially correlated, spectrally low-frequency, slowly varying along traverse) from residual EMI (less spatially correlated, broadband, with characteristic spectral features from equipment harmonics).
In validation on held-out field measurements, the combined pipeline reduces effective noise floor from the raw 150 to 200 pT/Hz to approximately 40 to 60 pT/Hz equivalent field noise after processing. This is the 50 pT/Hz sensitivity figure we quote in the product specification, and it is the processed output sensitivity, not the raw sensor sensitivity.
Where deconvolution has hard limits
The model performance degrades in conditions that differ significantly from the training distribution. The training dataset is weighted toward Western Australian and Queensland hard-rock terrains and specific types of industrial EMI from ventilation and haulage equipment. In environments with EMI sources not represented in training, such as DC rail traction return currents near electrified mine railways, or alkali-rich lithologies with magnetic susceptibility profiles outside the training range, the model's separation accuracy decreases.
We can measure this degradation by comparing model output against independent reference measurements in new environments, and we include this validation as part of any evaluation kit deployment. But we cannot promise training-set performance in out-of-distribution environments. A geophysicist relying on the 50 pT/Hz processed figure in an environment substantially different from the training distribution should treat that figure as an upper bound on processed sensitivity, not a guaranteed specification.
There is also a frequency range constraint. Below 0.5 Hz, both the reference channel subtraction and the sequence model become less effective because the geological gradient itself varies at those timescales as the instrument moves through the survey grid. Separating drift from gradient at sub-0.5-Hz timescales requires instrument velocity information that we do not currently integrate into the processing chain. This matters for very slowly traversed grids and for stationary measurements over extended intervals.
The inference architecture in practice
The sequence model runs on the embedded compute module in the instrument, not in cloud or post-processing software. Latency from raw spectrum acquisition to processed field estimate is under 120 milliseconds, which allows real-time display of the processed field value during field acquisition. The model is a compact recurrent architecture with 2.1 million parameters, quantised to 8-bit integer representation for the embedded target. We chose this architecture over transformer-based alternatives because the sequential nature of the magnetometry signal maps naturally to recurrent processing and the parameter count is manageable on the embedded hardware we use.
The processing pipeline is fully documented and can be run offline on raw data files if a customer wants to apply different processing parameters or compare outputs from different model versions. We do not black-box the processing output from customers who want to inspect the intermediate steps.
Distinguishing signal processing from sensor improvement
A point worth being explicit about: the deconvolution pipeline improves the effective sensitivity of the combined instrument system, but it does not change the physics of the NV-center sensor. The fundamental sensitivity floor, set by photon shot noise and spin coherence time, is a property of the diamond chip and the optical design. Processing cannot recover below that floor. Improving raw sensor sensitivity requires better diamond material, higher photon collection efficiency, or longer coherence times. That work runs in parallel with the signal processing development and is a separate engineering challenge.
The benefit of the processing pipeline is that it lets us operate closer to the physical sensitivity limit in real field conditions, where interference would otherwise mask the geological signal at the frequency range of interest. It is a practical tool, not a workaround for inadequate hardware.
The processing pipeline configuration can be adapted for specific EMI environments and geological settings. Contact the team to discuss what validation data we have for conditions similar to yours.
Contact the Team
