Multitask Learning Deep Neural Networks Enable Embedded Design of Active Metamaterials.

Area of Science:

• Photonics and optical engineering
• Materials science
• Artificial intelligence and machine learning

Background:

• Active metasurfaces offer tunable optical properties but their design is complex.
• Phase-change materials like Ge2Sb2Se4Te (GSST) enable dynamic control of optical characteristics.
• Efficient forward modeling and inverse design are crucial for realizing advanced photonic devices.

Purpose of the Study:

• To propose and implement a deep neural network framework for simplifying photonic device design.
• To develop a continuously tunable bandpass filter for the midwave infrared region.
• To demonstrate the effectiveness of multitask learning in optimizing devices with phase-change materials.

Main Methods:

• A deep neural network framework based on multitask learning was developed.
• The framework models crystallinity and geometric parameters of a metasurface-based Fabry-Perot cavity filter.
• The filter utilizes Ge2Sb2Se4Te (GSST) phase-change material controlled by a silicon heater.

Main Results:

• The deep learning framework successfully simplified the forward modeling and inverse design process.
• A continuously tunable bandpass filter operating in the midwave infrared was constructed and validated.
• The multitask deep learning strategy effectively maximized the use of GSST tuning for bandpass filtering.

Conclusions:

• Multitask deep learning provides a powerful tool for the on-demand reverse design of photonic structures.
• This approach enhances personalization and functional extension possibilities for optical devices.
• The study demonstrates a significant advancement in designing active photonic devices with phase-change materials.