Prediction multimodal optical responses for ultrafast plasmonic based functional universal approximation theorem with...

Area of Science:

• Nanophotonics
• Computational electromagnetics
• Machine learning

Background:

• Traditional machine learning methods for nanophotonics often struggle with functional data, requiring large datasets due to their inability to capture inherent data structures.
• Neural network (NN) approaches typically model functional data as high-dimensional vectors, neglecting smoothness and continuity, which limits their efficiency and robustness.

Purpose of the Study:

• To develop a data-efficient modeling framework for nanophotonic inverse design by leveraging functional data analysis (FDA).
• To address the limitations of traditional NN-based methods in capturing the functional nature of nanophotonic structure-optical response relationships.

Main Methods:

• Proposed a novel framework based on the functional universal approximation (FUA) theorem, integrating FDA principles.
• Explicitly modeled functional structure to learn nonlinear function-on-scalar mappings.
• Validated the FUA method on an aluminum nano ring-disk dimer model.

Main Results:

• Achieved high prediction accuracy for absorption (R ²=0.86), scattering (R ²=0.84), near-field spectra (R ²=0.96), and time-resolved electric fields (R ²=0.98).
• Demonstrated superior data efficiency, requiring only 300 training samples.
• Showcased enhanced robustness and generalization capabilities compared to standard NN models.

Conclusions:

• The FUA framework offers a powerful and data-efficient alternative for modeling nanophotonic systems.
• This approach shows significant promise for enabling efficient inverse design in nanophotonics, especially under limited data conditions.
• FUA effectively captures the functional nature of structure-property relationships, outperforming traditional NN methods in data efficiency and generalization.