Artificial neural network (ANN) is used to classify frequency resolved optical gating (FROG) spectrograms based on the phase of the corresponding pulse.

Generally, ultrafast spectroscopy requires a short pulse. When the different frequencies that consitute the pulse grow out of phase, the pulse grows longer and the time resolution of the spectroscopic measurement becomes longer. The pulse can be represented in the time domain or the frequency domain. The time domain shows us how short the pulse is and the frequency domain shows us if the different frequencies are in phase. A short, compressed pulse is shown below:

while an un-compressed pulse is shown here:

The pulse compression procedure consists of manipulating the phase of the uncompressed pulse until it resembles a compressed pulse. This can be achieved with a pulse shaper, such as an acousto-optic programmable dispserive filter (AOPDF). An AOPDF can apply a phase mask to correct the phase of an un-compressed pulse. The Dazzler AOPDF, by Fastlite, applies a fourth-order polynomial to the pulse phase. One then desirese to know what fourth-order polynomial best compresses a given pulse.

Frequency resolved optical gating (FROG) is a pulse characterization method that generates a spectrogram (x-axis: time, y-axis: frequency) that is defined by a given input pulse. The FROG corresponding a compressed pulse appears vertical, as there is no relative delay between frequencies:

However, the FROG from the un-compressed pulse is broader and has a characteristic shape that is defined by its spectral frequency:

The code provided in spectroFROG uses a 2D convolutional neural network to conduct multi-label regression on FROG spectrograms and returns three labels, each normalized from -1 to 1. The three labels correspond to the second-, third- and fourth-order polynomial coefficients of the spectral phase of the input pulse.

The neural network can then provide three polynomial phase labels for a FROG trace corresponding to an unkown pulse.

The model training converges to an accuracy of 94% after 20 epochs: