Sub-photon accuracy noise reduction of a single shot coherent diffraction pattern with an atomic model trained...

Area of Science:

• * Physics
• * Materials Science
• * Data Science

Background:

• * Single-shot imaging with femtosecond X-ray lasers offers high spatial and temporal resolution.
• * Conventional noise reduction methods are challenging for this technique, limiting accuracy.
• * Deep learning approaches show promise for overcoming these limitations.

Purpose of the Study:

• * To validate deep learning-based noise reduction techniques for femtosecond X-ray laser imaging.
• * To investigate the performance of autoencoder neural network architectures for this application.
• * To assess the transfer learning capabilities of these models for diverse nanoparticle imaging.

Main Methods:

• * Simulated a large dataset of nanoparticle diffraction patterns, modeling nanoparticles as collections of atoms.
• * Investigated three neural network architectures: neural network, convolutional neural network, and U-net.
• * Applied the best-performing U-net model to experimental coherent diffractive imaging data of a nanoparticle in a microfluidic device.

Main Results:

• * The U-net architecture demonstrated superior performance in noise reduction and sub-photon reproduction compared to other models.
• * The trained U-net model successfully reduced noise in experimental X-ray laser diffraction patterns.
• * Reconstructed nanoparticle images showed significant improvement in quality and detail after noise reduction.

Conclusions:

• * Deep learning, particularly U-net, is effective for noise reduction in single-shot femtosecond X-ray laser imaging.
• * The developed models exhibit strong transfer learning capabilities, enabling accurate imaging of novel nanoparticle shapes.
• * This approach significantly enhances the accuracy and applicability of coherent diffractive imaging.