Noise reduction in X-ray photon correlation spectroscopy with convolutional neural networks encoder-decoder models.

Area of Science:

• Materials Science
• Spectroscopy
• Computational Physics

Background:

• X-ray photon correlation spectroscopy (XPCS) is vital for studying sample dynamics but is susceptible to noise.
• Noise sources include random/correlated fluctuations and heterogeneities, obscuring intrinsic sample dynamics in correlation functions.
• Addressing multiple noise origins simultaneously in experimental data is a significant challenge.

Purpose of the Study:

• To introduce a novel computational method for enhancing signal-to-noise ratio in XPCS.
• To leverage convolutional neural network encoder-decoder (CNN-ED) models for noise reduction in two-time correlation functions.
• To improve the quantitative analysis of sample dynamics from noisy experimental data.

Main Methods:

• Utilized convolutional neural network encoder-decoder (CNN-ED) models for signal processing.
• Trained CNN-ED models on real-world experimental XPCS data.
• Employed convolutional and transposed convolutional layers to extract features and reconstruct clean correlation functions.

Main Results:

• Demonstrated effective noise reduction in two-time correlation functions using CNN-ED models.
• Successfully extracted equilibrium dynamics' parameters from data containing statistical noise and dynamic heterogeneities.
• Showcased the ability of CNN-ED models to learn signal functional forms for enhanced quantitative analysis.

Conclusions:

• CNN-ED models offer a powerful computational approach to improve signal quality in XPCS.
• This method enhances the quantitative utility of low signal-to-noise XPCS data.
• Further strategies for optimizing model performance and defining applicability limits were discussed.