Generative Adversarial Neural Networks for Denoising Coherent Multidimensional Spectra.

Abstract

Ultrafast spectroscopy often involves measuring weak signals and long data acquisition times. Spectra are typically collected as a "pump-probe" spectrum by measuring differences in intensity across laser shots. Shot-to-shot intensity fluctuations are most often the primary source of noise in ultrafast spectroscopy. Here, we present a novel approach for denoising ultrafast two-dimensional infrared (2D IR) spectra using conditional generative adversarial neural networks (cGANNs). The cGANN approach is able to eliminate shot-to-shot noise and reconstruct the line shapes present in the noisy input spectrum. We present a general approach for training the cGANN using matched pairs of noisy and clean synthetic 2D IR spectra based on the Kubo-line shape model for a three-level system. Experimental shot-to-shot laser noise is added to synthetic spectra to recreate the noise profile present in measured experimental spectra. The cGANNs can recover line shapes from synthetic 2D IR spectra with signal-to-noise ratios as low as 2:1, while largely preserving the key features such as center frequencies, line widths, and diagonal elongation. In addition, we benchmark the performance of the cGANN using experimental 2D IR spectra of an ester carbonyl vibrational probe and demonstrate that, by applying the cGANN denoising approach, we can extract the frequency-frequency time correlation function (FFCF) from reconstructed spectra using a nodal-line slope analysis. Finally, we provide a set of practical guidelines for extending the denoising method to other coherent multidimensional spectroscopies.