General deep learning framework for emissivity engineering.

Area of Science:

• Metamaterials and Nanophotonics
• Artificial Intelligence in Materials Science

Background:

• Wavelength-selective thermal emitters (WS-TEs) are crucial for applications like thermal camouflage, radiative cooling, and gas sensing.
• Existing design methods lack a general framework, requiring application-specific materials and structures, and often fail to co-optimize both.
• Previous approaches typically fix either material composition or structural parameters, limiting design flexibility.

Purpose of the Study:

• To establish a general design framework for emissivity engineering of WS-TEs.
• To develop an AI-driven approach for autonomous material selection and structural optimization.
• To demonstrate the framework's versatility across diverse applications.

Main Methods:

• Employed a deep Q-learning network, a reinforcement learning algorithm, for designing multilayer WS-TEs.
• Created a self-built material library for autonomous material selection.
• Optimized structural parameters to achieve target emissivity spectra.

Main Results:

• Successfully designed and fabricated three distinct WS-TEs for thermal camouflage, radiative cooling, and gas sensing.
• Demonstrated the deep Q-learning algorithm's ability to offer a general design framework beyond 1D multilayer structures.
• Validated the autonomous selection of materials and optimization of structural parameters for target spectral emissivity.

Conclusions:

• The deep Q-learning framework provides a feasible and efficient approach for designing WS-TEs across various applications.
• The framework's scalability in materials, structures, and dimensions offers a generalized solution for emissivity engineering.
• This work paves the way for efficient design in nonlinear optimization problems, extending beyond thermal metamaterials.