RF-spectrum-based noise characterization of pulsed lasers using artificial intelligence models

Ultrashort pulsed lasers are extensively employed in engineering and physical sciences, with many applications requiring repetition rate and power fluctuation stabilization. Currently, the scientific literature offers high-precision methods for noise characterization. However, they often require complex optical setups. In this study, we demonstrate a convolutional neural network (CNN), a Kolmogorov–Arnold network (KAN), and a symbolic regression (SR) that are able to predict amplitude and timing noise using the radio frequency power spectrum of a mode-locked pulsed laser. The CNN, KAN, and SR predictions are compared with each other and with a theoretical model well known in the scientific literature. Fluctuations in the pulse amplitude and pulse repetition time can be readily recognized with high precision from their characteristic spectral signature using these artificial intelligence models. It is demonstrated that the CNN, KAN, and SR outperform a theoretical model well known in the scientific literature. Among the four approaches investigated in this work, KAN and SR presented superior prediction capabilities for amplitude and time jitter with an R 2 =0.99 for all cases considered (training, validation, and test sets). Specifically, for amplitude jitter, the KAN and SR presented RMSE values of 0.1%, while for time jitter, the RMSE values were 16 and 4 ps, for KAN and SR, respectively.