What is it about?
Ultrafast laser pulse shaping requires controlling the spectral phase of femtosecond pulses with high precision. The standard approach uses adaptive feedback loops: measure the pulse, adjust the shaper, repeat. This works but converges slowly and does not generalize -- the learned corrections apply only to the specific hardware configuration used during training. We applied semi-supervised learning to this problem. The system first learns a forward model of the pulse shaper (predicting output pulse shape from shaper settings) from a small labeled dataset, then refines it using a larger set of unlabeled measurements. The learned model can then be inverted to compute shaper settings for a desired target pulse. The semi-supervised approach reduces the amount of labeled training data needed and produces corrections that transfer better across hardware configurations than purely supervised methods.
Featured Image
Photo by Gritte on Unsplash
Why is it important?
Pulse shaping hardware has physical imperfections (nonlinear responses, cross-talk between spectral channels) that limit the fidelity of shaped pulses. Traditional feedback loops correct for these imperfections implicitly but do not build a transferable model of the device. The semi-supervised approach builds an explicit model that separates the device physics from the target pulse shape. This means corrections learned on one set of target pulses generalize to new targets without retraining from scratch. The reduction in labeled data requirements also matters practically, since acquiring labeled pulse measurements is time-consuming on the experimental side.
Perspectives
This paper came from working across two positions -- one in HPC for chemical engineering, one as a project associate in optics. The pulse shaping work was where I first got interested in combining machine learning with experimental physics. The semi-supervised framing arose from a practical constraint: labeled data (measured pulse shapes paired with shaper settings) was expensive to collect, but unlabeled shaper responses were cheap. The asymmetry mapped naturally onto semi-supervised learning. The results were promising but preliminary. The approach pointed toward learning-based control of experimental hardware, which has since become a much more active research area.
Rohit Goswami
University of Iceland
Read the Original
This page is a summary of: Semi-Supervised Approaches to Ultrafast Pulse Shaping, January 2021, Springer Science + Business Media,
DOI: 10.1007/978-981-15-9259-1_172.
You can read the full text:
Contributors
The following have contributed to this page
