Deep learning-based single-shot computational spectrometer using multilayer thin films.

Area of Science:

• Optics and Photonics
• Spectroscopy
• Machine Learning Applications

Background:

• Computational spectrometers offer potential for mobile applications due to their size and speed.
• Previous demonstrations were limited to simple spectra, hindering broader applications.

Purpose of the Study:

• To develop a deep learning (DL)-based single-shot computational spectrometer.
• To enable reconstruction of both narrow and broad spectra for mobile sensing.

Main Methods:

• Fabrication of a multilayer thin-film filter array using wafer-level stencil lithography.
• Integration of the filter array with a CMOS image sensor for single-exposure image capture.
• Development of a DL architecture (dense layer, U-Net backbone with residual connections) for spectrum reconstruction.

Main Results:

• Accurate reconstruction of 323 test spectra (500-850 nm, 1 nm spacing) with a low root mean squared error of 0.0288.
• Validation of multilayer thin-film filters via SEM, confirming uniform deposition and high yield.
• Demonstration of a compact, fast, and accurate computational spectrometer compatible with CMOS sensors.

Conclusions:

• The developed DL-based computational spectrometer effectively reconstructs complex spectra from single-shot measurements.
• The device's performance and features make it suitable for commercialization in mobile applications.
• This technology advances portable spectroscopic analysis for on-site detection and self-diagnosis.