Diamonds contain tiny defects: instead of two carbon atoms, there's a nitrogen atom and a vacancy. Such a defect senses magnetic fields like a compass needle senses north. When the field changes, the defect changes its emission frequency. But to turn that emission into numbers, the sensor needs lengthy calibration against a reference.

Researchers proposed teaching the algorithm not with labeled examples, but with the law of nature itself. They embedded the rule by which a magnetic field splits the defect's energy levels into the learning program. Now the algorithm doesn't need thousands of examples with known fields: it understands the signal's structure on its own. As a result, the measurement error dropped by a factor of 372 compared to the standard statistical approach.

The method was tested on a real diamond sensor without prior calibration. The algorithm took raw data and produced an accurate magnetic field value. This is a step toward sensors that don't need calibration before every use. The approach could also benefit other rare physical measurements where data is scarce but the laws are known.