Deep-learning-assisted identification of sub-diffraction nanodomains in LiNbO₃ crystals

Area of Science:

• Materials Science
• Nanotechnology
• Optics

Background:

• Lithium niobate (LiNbO₃) domain structures are crucial for nonlinear optics, quantum applications, and ferroelectric memory.
• Advances in nanoscale engineering enable sub-diffraction LiNbO₃ nanodomains for high-speed modulation and storage.
• Characterizing these nanoscale domains remains a significant challenge due to resolution limitations.

Purpose of the Study:

• To propose and demonstrate a deep-learning-assisted method for identifying sub-diffraction LiNbO₃ nanodomain lines.
• To overcome the spatial resolution limits of conventional imaging techniques for nanodomain characterization.
• To provide an efficient and feasible approach for analyzing nanoscale domain structures.

Main Methods:

• Experimental recording of second-harmonic (SH) images of LiNbO₃ nanodomain lines using a confocal microscope.
• Fabrication of nanodomains with linewidths ranging from 200 nm to 600 nm, below the microscope's resolution (∼800 nm).
• Training a neural network with 1568 SH images to recognize and differentiate nanodomain lines.

Main Results:

• The deep-learning model achieved an 81.25% accuracy in recognizing different nanodomain lines.
• The method successfully identified sub-diffraction nanodomains from images limited by optical resolution.
• Demonstrated the feasibility of using deep learning for nanoscale domain characterization.

Conclusions:

• Deep learning offers an efficient solution for identifying sub-diffraction LiNbO₃ nanodomains.
• This approach enhances the characterization capabilities for advanced LiNbO₃-based devices.
• The study paves the way for improved high-speed modulation and high-capacity data storage applications.