Dynamic coherent diffractive imaging with a physics-driven untrained learning method.

Area of Science:

• Computational imaging
• Coherent diffraction imaging (CDI)

Background:

• Reconstructing complex fields from single diffraction measurements is challenging.
• Conventional iterative algorithms are slow and struggle with convergence.
• Deep learning methods require large, often unobtainable, training datasets.

Purpose of the Study:

• To introduce a physics-driven, untrained learning method for complex field reconstruction.
• To enable high-confidence and fast imaging of dynamic processes using CDI.
• To overcome limitations of conventional and deep learning approaches in CDI.

Main Methods:

• Developed Deep CDI, combining a neural network with a physical imaging model.
• Utilized support region and free propagation constraints for loss calculation.
• Optimized network weights by summing losses from both constraints.

Main Results:

• Demonstrated feasibility through numerical simulations and optical experiments on static samples.
• Successfully imaged a dynamic process using 3600 diffraction patterns.
• Achieved an average reconstruction speed of 228 frames per second.

Conclusions:

• Deep CDI offers a viable solution for complex field reconstruction from single diffraction patterns.
• The method effectively images dynamic processes with high speed and confidence.
• Physics-driven untrained learning provides an alternative to data-intensive deep learning methods in CDI.