Misalignment resilient phase-filtered diffractive deep neural networks.

Area of Science:

• Optics
• Artificial Intelligence
• Photonics

Background:

• Diffractive deep neural networks (D2NN) are effective for wavefront shaping and beam manipulation.
• High-precision alignment across multiple planes is a critical challenge for D2NNs.

Purpose of the Study:

• To propose a phase-filtered diffractive deep neural network (PF-D2NN) to enhance robustness against alignment errors.
• To improve the reliability of D2NNs in optical systems.

Main Methods:

• Introduced a phase filtering operator during phase optimization for modulation layers.
• Developed a back-propagation (BP) algorithm tailored for PF-D2NN phase optimization.
• Validated the approach through both simulations and experimental setups.

Main Results:

• The PF-D2NN demonstrated significant robustness to alignment errors compared to conventional D2NNs.
• Clear images were consistently produced by PF-D2NN even with misalignments exceeding 5 pixels.
• Conventional D2NNs failed to produce clear images under similar misalignment conditions.

Conclusions:

• The proposed PF-D2NN effectively overcomes the alignment sensitivity limitations of traditional D2NNs.
• PF-D2NN offers a more reliable solution for wavefront shaping and beam manipulation in practical applications.
• This advancement paves the way for more robust optical AI systems.