Deep empirical neural network for optical phase retrieval over a scattering medium.

Area of Science:

• Physics
• Machine Learning
• Information Science

Background:

• Supervised learning requires extensive labeled data, limiting its application in complex scientific domains.
• Physics-enhanced deep neural networks incorporate analytical models but fail for systems lacking analytical solutions, such as multi-input/multi-output wave scattering systems.

Purpose of the Study:

• To introduce a novel deep empirical neural network (DENN) approach for analyzing systems without analytical solutions.
• To demonstrate DENN's capability in 'seeing through' opaque scattering media in an untrained manner.

Main Methods:

• Developed a hybrid deep neural network and empirical model architecture (DENN).
• Applied DENN to optical phase retrieval problems, specifically in opaque scattering media.
• Evaluated DENN's performance against supervised learning methods without relying on labeled data.

Main Results:

• DENN achieved a 58% improvement in fidelity for optical phase retrieval compared to supervised learning.
• The DENN approach successfully operated without requiring labeled training data.
• Demonstrated the ability to analyze complex wave scattering systems.

Conclusions:

• Deep empirical neural networks (DENN) provide an effective, data-efficient solution for scientific problems lacking analytical models.
• DENN shows significant potential for advancing deep learning applications across physics, information science, biology, and chemistry.