Enhancing All-Optical NV Center Magnetometry With Machine Learning: Model Selection for Efficient and Deployable Quantum Sensors

Negatively charged nitrogen–vacancy (NV) cen- ters in diamond are promising platforms for magnetic field sensing under ambient conditions, yet practical deployment is often limited by hardware complexity, power consumption, and computational cost. All-optical approaches based on fluorescence lifetime measurements avoid microwave exci- tation and enable simpler, potentially more scalable sensor designs, but the nonlinear response to magnetic and thermal fields requires advanced modeling. In this study, a systematic comparison of 18 regression algorithms—including linear models, ensemble methods, support vector machines, and symbolic regression—was conducted for predicting mag- netic field strength from NV fluorescence lifetime data. A dataset of 22 758 measurements was employed, and models were assessed in terms of both predictive accuracy and computational footprint (training time, model complexity, and inference latency). The results indicate that ensemble techniques such as LightGBM and XGBoost achieve near-perfect accuracy while maintaining subsecond inference times, making them suitable for real-time embedded deployment. In contrast, symbolic regression provides negligible inference cost but suffers from significantly lower accuracy. These findings offer quantitative guidance for the selection of machine learning models that balance precision and efficiency, paving the way toward compact, low-power, all-optical NV magnetometers for resource-constrained platforms.