Non-Contact Thermometry of Silicon via Broadband Supercontinuum and Machine Learning

A new approach to non-contact picosecond thermometry in silicon, which combines broadband supercontinuum spectroscopy with machine learning, is presented. The technique reconstructs temperature from absorption spectra with a mean absolute error below 0.8% and sub-nanosecond temporal resolution. Comprehensive evaluation of neural network architectures identifies an optimal configuration featuring an initial one-dimensional convolutional layer followed by residual blocks, which substantially outperforms the classical approximation approach and fully connected networks. The model demonstrates robustness to different thermal distributions, maintaining accuracy under both uniform and laser heating. Spectral sensitivity analysis confirms the network’s reliance on physically meaningful spectral features near silicon’s bandgap (1100–1200 nm). Although extrapolation beyond the training temperature range remains challenging, this method establishes a powerful framework for ultrafast thermal characterization of semiconductors with broad applications in photonics and laser material processing.