Machine Learning Based Optical Thermometry Using Photoluminescence and Raman Spectra of Diamonds Containing SiV Centers

Abstract

Micro- and nanothermometry enable precise temperature monitoring and control at the micro- and nanoscale, and have become essential diagnostic tools in applications ranging from high-power microelectronics to biosensing and nanomedicine. Most existing techniques rely on secondary micro- and nanothermometers that require individual calibration of each sensor, ideally both off- and in-situ, before use. We present an alternative approach that overcomes this limitation by employing fluorescent diamonds containing silicon-vacancy centers, where the thermo-sensitive physical quantities are the centers' photoluminescence and the diamond host's Raman signals. The photoluminescence and Raman data are analyzed using two multi-feature regression algorithms that leverage a minimal number of calibration diamonds and temperature set points to predict the temperature of previously unseen diamonds. Using this approach, the models achieve accuracies as low as 0.7 K, resolutions down to 0.6 K Hz$^{-1/2}$, and sensitivity as high as 0.04 K$^{-1}$. These correspond to improvements of roughly 70 percent (over threefold) in accuracy, 50 percent (twofold) in resolution, and 567 percent (sevenfold) in sensitivity compared with traditional single-feature models. Our approach is particularly suited to applications where pre-deployment calibration of every thermosensor is impractical, and it is generalizable to any thermometry platform with two or more simultaneously measurable temperature-dependent observables.