Terahertz ptychography enabled by untrained physics-driven neural networks.

Area of Science:

• Physics
• Imaging Science
• Computational Science

Background:

• Ptychography is a phase imaging technique using translated intensity images.
• Deep learning shows promise for phase retrieval but requires large labeled datasets.
• Acquiring terahertz (THz) labeled data for neural network pretraining is challenging, limiting generalization.

Purpose of the Study:

• To develop a high-quality reconstruction method for THz ptychography.
• To overcome the data limitations in deep learning for THz imaging.
• To introduce an adaptable and efficient phase retrieval technique.

Main Methods:

• Proposed an untrained physics-driven neural network (UPNN).
• UPNN integrates real-world physical models to guide neural network inference.
• The network requires only a single dataset, eliminating the need for pretraining.

Main Results:

• UPNN achieves high-quality reconstruction of THz ptychography.
• Demonstrated superior performance over conventional iterative ptychography algorithms.
• Showcased improvements in image quality, lateral resolution, and robustness.

Conclusions:

• UPNN offers a flexible and effective solution for THz ptychography reconstruction.
• The method overcomes the data acquisition challenges in deep learning for THz imaging.
• UPNN provides a robust alternative to traditional algorithms for phase retrieval.