Stochastic minibatch approach to the ptychographic iterative engine.

Area of Science:

• X-ray microscopy and imaging science.
• Computational imaging and phase retrieval algorithms.

Background:

• The ptychographic iterative engine (PIE) is a key algorithm for nanometer-scale phase retrieval in imaging experiments.
• PIE's optimization performance is sensitive to the order of diffraction intensity data, often leading to stagnation in local solutions.
• Current random ordering methods limit PIE's ability to escape suboptimal results.

Purpose of the Study:

• To introduce an improved PIE algorithm by incorporating machine learning training concepts.
• To enhance the convergence properties and efficiency of the phase retrieval process in PIE.
• To overcome limitations associated with traditional random ordering in PIE.

Main Methods:

• Extension of the ptychographic iterative engine (PIE) algorithm.
• Application of minibatch stochastic gradient descent, a machine learning training technique.
• Analysis of diffraction intensities from multiple scanning locations with varying data ordering.

Main Results:

• Demonstrated significant improvements in the convergence properties of the PIE numerical optimization.
• The new machine learning-based approach effectively addresses local solution stagnation.
• Enhanced efficiency and reliability in achieving high-resolution phase retrieval.

Conclusions:

• The integration of minibatch stochastic gradient descent substantially improves PIE performance.
• This advancement offers a more robust method for nanometer-scale phase retrieval.
• The findings pave the way for more effective computational imaging techniques.