Machine learning-enabled synchrotron FTIR thermometry for thermal mapping in nanocalorimetry chips

Accurate temperature mapping at the microscale is essential for the reliable operation of nanocalorimetric devices, particularly under synchrotron-based conditions where localized heating and steep thermal gradients can significantly influence experimental outcomes. Here, we present a non-contact and data-driven approach that combines synchrotron radiation-based Fourier-transform infrared (SR-FTIR) thermometry with machine learning (ML) to reconstruct spatial temperature distributions in SiN x -SiO x nanocalorimetry chips. Temperature-dependent infrared spectra were used to train multiple regression models, including linear, ensemble, kernel-based, and deep learning approaches. Among the evaluated models, the convolutional neural network achieved the lowest global prediction error (with a coefficient of determination R 2 = 0.997 and root mean square error, RMSE = 6.53 ºC), whereas Gaussian process regression provided the most robust performance for spatial thermometry, combining high accuracy (R 2 = 0.996, RMSE = 7.10 ºC) with the lowest residual standard deviation between predicted and actual temperatures (3.5 ºC, compared with 7.3 ºC for CNN) and smoother residual behaviour. Model interpretability was addressed using SHAP analysis, revealing that the most influential spectral regions coincide with thermally responsive vibrational modes of the Si-O network. Finite element modelling (FEM), incorporating experimentally determined emissivity and convective boundary conditions, was used to simulate heat transport and validate the ML-derived temperature maps. The good agreement between FEM and ML predictions confirms the reliability of the approach while highlighting the ability of ML to capture complex experimental effects not explicitly included in the physical model. This integrated framework enables data-driven and spatially resolved thermal diagnostics in microscale systems and is readily extendable to other material platforms and operando synchrotron experiments.