Speaker
Description
Accurate microscale temperature mapping is critical for the reliable operation of nanocalorimetric devices, particularly under synchrotron-based experimental conditions, where localized heating and steep thermal gradients can strongly affect measurement outcomes. In this work, we present a non-contact, data-driven methodology that combines synchrotron radiation Fourier-transform infrared spectroscopy (SR-FTIR) thermometry with machine learning to reconstruct spatial temperature distributions in SiNₓ–SiOₓ nanocalorimetry chips. Temperature-dependent infrared spectra were used to train and evaluate several regression models, including linear, ensemble, kernel-based, and deep learning approaches. Among these, Gaussian Process Regression achieved the best performance, with R² = 0.996 and RMSE = 7.1 °C, demonstrating high predictive accuracy and robust generalization across experimental conditions. Model interpretability was assessed using SHAP analysis, which identified the most relevant spectral regions as those associated with thermally responsive vibrational modes of the Si–O network. Finite element modelling, incorporating experimentally determined emissivity and convective boundary conditions, was further employed to simulate heat transport and validate the machine-learning-derived temperature maps. The close agreement between finite element simulations and machine learning predictions confirms the reliability of the proposed approach, while also highlighting the capacity of machine learning to capture complex experimental effects not explicitly represented in the physical model. Overall, this integrated framework provides a rapid, non-contact, and spatially resolved strategy for thermal diagnostics in microscale systems, with potential applicability to other material platforms and operando synchrotron experiments.