Weak signal extraction enabled by deep neural network denoising of diffraction data.

Area of Science:

• Materials Science
• Data Science
• Physics

Background:

• Denoising is crucial for scientific data, demanding accurate reproduction of ground truth.
• Existing denoising methods struggle with unknown and multiple noise sources common in scientific data.
• Simulation-based denoising is limited by the complexity of real-world noise profiles.

Purpose of the Study:

• To develop a robust denoising strategy for scientific data, particularly X-ray diffraction and scattering data.
• To enable the accurate visualization and analysis of weak signals obscured by noise.
• To establish a practical noise filtering approach for challenging data acquisition scenarios.

Main Methods:

• Supervised training of a deep convolutional neural network (CNN).
• Utilizing pairs of measured low- and high-noise data for network training.
• Applying the trained CNN to denoise X-ray diffraction and resonant X-ray scattering data from crystalline materials.

Main Results:

• Weak signals, such as those from charge ordering, become visible and quantitatively accurate after denoising.
• The CNN successfully recovers subtle signals that are insignificant in the original noisy data.
• Training with artificially generated noise did not achieve the same level of quantitative accuracy.

Conclusions:

• Deep convolutional neural networks offer a practical and effective solution for denoising scientific data.
• Supervised learning with real noisy data pairs is essential for achieving quantitative accuracy in signal recovery.
• This approach enhances the applicability of techniques like X-ray scattering for studying subtle material properties.