Inverse design of core-shell particles with discrete material classes using neural networks

Area of Science:

• Materials Science
• Computational Physics
• Optical Engineering

Background:

• Designing particles with specific scattering properties is complex.
• Current methods for designing optical scatterers are often time-consuming and inflexible.

Purpose of the Study:

• To develop a machine learning-assisted framework for designing multi-layered core-shell particles with on-demand scattering responses.
• To create a faster and more flexible approach for inverse design of optical scatterers.

Main Methods:

• Utilized artificial neural networks (ANNs) as differentiable surrogate models for gradient-based optimization.
• Implemented a two-step optimization process to handle continuous geometric parameters and discrete material choices simultaneously.
• Employed a classification network to manage varying numbers of shells, addressing problem non-uniqueness and expanding design space.

Main Results:

• Achieved high accuracy in predicting scattering spectra using ANNs.
• Demonstrated a method that is 1-2 orders of magnitude faster than conventional approaches for both forward prediction and inverse design.
• Successfully designed core-shell particles with tailored scattering responses, considering material constraints and variable shell numbers.

Conclusions:

• The proposed machine learning framework offers a significant speedup and enhanced flexibility for designing custom optical scatterers.
• This approach is scalable for larger and more complex scatterer designs.
• The method facilitates the on-demand design of multi-layered core-shell particles with precise scattering characteristics.