Cross-domain heterogeneous metasurface inverse design based on a transfer learning method

Area of Science:

• Metasurface design
• Computational electromagnetics
• Machine learning applications

Background:

• Metasurface design often requires large datasets for inverse design.
• Heterogeneous metasurface datasets with distinct functionalities pose challenges for model training.
• Transfer learning offers a potential solution to mitigate data scarcity in scientific domains.

Purpose of the Study:

• To propose a transfer learning method for metasurface inverse design.
• To enable knowledge transfer across heterogeneous metasurface datasets with diverse functionalities.
• To reduce the data threshold for training metasurface inverse design models.

Main Methods:

• Utilized a transfer learning approach by fine-tuning an inverse design neural network.
• Froze the parameters of hidden layers to preserve learned features.
• Transferred knowledge from an electromagnetic-induced transparency (EIT) domain to absorption and phase-controlled metasurface domains.

Main Results:

• Successfully transferred metasurface inverse design knowledge to new domains.
• Achieved training with a minimum of 700 target domain samples, significantly lower than source domain requirements.
• Demonstrated effective knowledge transfer for EIT (different design), absorption, and phase-controlled metasurfaces.

Conclusions:

• The proposed transfer learning method effectively lowers the data threshold for metasurface inverse design.
• This work highlights the potential for knowledge transfer between different metasurface datasets and functionalities.
• The findings pave the way for more efficient and data-driven metasurface design.