Deep Convolutional Mixture Density Network for Inverse Design of Layered Photonic Structures

Area of Science:

• Nanophotonics
• Machine Learning
• Materials Science

Background:

• Machine learning (ML) is crucial for nanophotonic inverse design.
• Nonuniqueness in optical spectra hinders conventional ML algorithm convergence.
• Vastly different designs can yield identical spectral properties, posing a challenge.

Purpose of the Study:

• To address the nonuniqueness problem in ML-driven nanophotonic inverse design.
• To develop an ML approach capable of handling multimodal probability distributions for design parameters.
• To enable convergence for non-unique solutions without losing degenerate possibilities.

Main Methods:

• Introduced a mixture density network (MDN) approach.
• Modeled design parameters as multimodal probability distributions instead of discrete values.
• Applied the MDN technique to inversely design multilayer oxide thin-film photonic structures (10-layer and 4-layer cases).

Main Results:

• The MDN successfully handled complex transmission spectra and varying illumination conditions for a 10-layer structure.
• The 4-layer case demonstrated stronger multimodal characteristics, revealing alternative spectral solutions.
• The probabilistic output of the MDN provides valuable postprocessing information on prediction uncertainty.

Conclusions:

• The MDN approach offers an effective solution for the inverse design of photonic structures, particularly those with high degeneracy and spectral complexity.
• This method enhances the search for optimal nanophotonic designs by accommodating non-unique solutions.
• The technique provides insights into design uncertainties, improving the reliability of ML-driven inverse design.