Physics-informed neural operators for ultrafast spectroscopy data reconstruction
Abstract
Physics-informed neural operators (PINO) combine deep learning with mesh invariance and differentiable physics to make it easier to reconstruct ultrafast spectroscopic data. This method lets scientists see excited states that only last a very short time, usually from pico- to femtoseconds. But getting great temporal resolution is hard because of the limits of the technology. Data gathering frequently leads to a diminished collection of photons at each frequency-time coordinate, hence constraining phase sweeps and complicating reconstruction endeavours. It is very important to accurately reconstruct time-domain data from a small number of frequency-time measurements since this is an ill-posed inverse issue that impacts the quality and reliability of the data. In this context, using physics-informed machine learning techniques is a promising way to improve inference and reconstruction, which will help ultrafast spectroscopy analysis move further.