Diamond thermometry beyond the zero-phonon line via physics-aware spectral trend learning

Nitrogen-vacancy centers in diamond enable microwave-free optical thermometry, but conventional zero-phonon-line readout becomes unreliable at elevated temperature and low signal-to-noise ratios because the local spectral signature progressively weakens into the background. Here, we infer temperature from the thermodynamic evolution of the phonon sideband (PSB) across a broad spectral window. A physics-aware decomposition separates a curvature-preserving PSB trend, which retains thermally driven broadening and redshift, from an ultra-smooth baseline that captures device-dependent background drift. These descriptors form a five-channel input to a one-dimensional dilated residual network with coarse-to-fine regression, shifting temperature inference from fragile local peak tracking to robust global trend learning. In the enhanced-inference noise-stress diagnostic, the selected σtrain=0.04 configuration gives the best clean-test mean absolute error (MAE) of 0.21 K,while the full σtrain sweep characterizes the trade-off between clean-condition accuracy and severe-noise robustness. A strict exact-match single-pass benchmark, disabling inference-time enhancement and post-hoc calibration, gives a five-channel ResNet MAE of 0.6792 K and confirms that the gain arises from the physics-aware representation–backbone combination rather than from inference enhancement alone. Beyond the in-domain benchmark, the pretrained representation also supports lightweight target-domain transfer: on a related sample batch over a narrower 20–60 °C interval, few-step continuation reduces the all-sample diagnostic MAE from 0.484K to 0.045 K. This transfer result demonstrates efficient target-domain refinement, but is reported separately from the main in-domain benchmark owing to the different temperature interval and sample domain. These results establish PSB-wide physics-aware trend learning as a robust route for microwave-free diamond thermometry and practical raster-scanned thermal mapping.