Inside a piece of diamond hides a tiny defect — a nitrogen atom next to a vacancy. This defect catches magnetic fields like an antenna catches radio waves and responds with flashes of light. Such sensors already measure temperature inside living cells and find defects in microchips.

The trouble is that the signal from a single defect drowns in noise. To extract the useful part, the instrument repeats the measurement hundreds of times and averages. It's like listening to a whisper on a noisy street: you need to ask the person to repeat the phrase many times to finally make out the words.

The neural network NVRNet does it differently. It was trained on computer models of how the defect responds to magnetic fields. Then it was fine-tuned on real data from three different defects. Now it takes one or two noisy signals and outputs a clean wave, along with the parameters of neighboring carbon atoms. The error turned out to be half that of the raw signal.

This approach speeds up measurements by up to 40 times. Instead of hours — minutes. This paves the way for fast tuning of quantum devices and mass scanning of materials.