A Fully Convolutional Approach to Denoising Material Dynamics Data from X-Ray Photon Correlation Spectroscopy

X-ray photon correlation spectroscopy (XPCS) is a technique used to measure the dynamics of materials through coherent X-ray beams. The nonequilibrium dynamics of the sample are studied using a two-time correlation function (2TCF), which is evaluated from the beam scattered off the sample. The 2TCF is subject to various types of noise, including random and correlated fluctuations, as well as heterogeneities that obscure the average dynamic parameters of the sample. The correlated nature of the noise makes off-the-shelf solutions, such as Gaussian filters, inadequate for denoising such data.<br/>Previously reported deep learning models based on a convolutional neural network encoder-decoder (CNN-ED) architecture have successfully improved the signal-to-noise ratio in the 2TCF. However, these models are limited to specific dynamics (e.g., equilibrium) with a fixed number of time samples (lags). A newly developed fully convolutional model can be applied to denoise a wide range of material dynamics and lags, eliminating the need for multiple customized models. In earlier models, a linear latent space layer played a crucial role in denoising by filtering out unwanted features from the input. The proposed model, which consists solely of convolutional layers without a latent space, can learn the functional form of the signal through supervised training. We demonstrate that fully convolutional encoder-decoder models trained on experimental data can effectively suppress noise in the 2TCF for a variety of material dynamics at NSLS-II. Noise reduction enhances quantitative usage of XPCS data and creates the potential for automating the analytical workflow, which is key to autonomous experiments.