In particle scattering experiments, the target is often polarized: nuclei are aligned so that their spins point predominantly in one direction. The degree of this alignment affects the collision outcomes. To measure it, a weak radio signal is sent through the target, and the absorption by the nuclei is monitored. It's like tuning a radio: the useful signal is drowned in noise, and its volume fluctuates due to interference.

Traditionally, the signal is extracted using phase-sensitive detection — essentially a lock-in amplifier that listens only to the desired frequency. But even that doesn't save you from noise, baseline drift, and systematic errors. In this new work, a neural network was applied for the first time to learn how to separate signal from noise using many examples. The network was trained on realistic simulations and real data, and it was able to significantly reduce the curve-fitting error.

The improvement in accuracy directly impacts experimental quality: the smaller the polarization uncertainty, the more reliable the conclusions about particle interactions. The method is suitable for real-time online monitoring and for subsequent data analysis. It's a new tool for NMR polarimetry in high-energy physics.