Exploiting geometric biases in inverse nano-optical problems using artificial neural networks

Area of Science:

• Nano-optics
• Computational electromagnetics
• Artificial Intelligence

Background:

• Solving inverse problems in nano-optics is challenging, often requiring solutions that meet specific constraints.
• Formulating prior information to select plausible solutions for nano-optical inverse problems remains unclear.

Purpose of the Study:

• To investigate the use of artificial neural networks (ANNs) for solving nano-optical inverse problems.
• To develop a method for reconstructing object shapes using electromagnetic field data and prior assumptions.
• To compare the performance of ANNs, topology optimization, and a hybrid approach.

Main Methods:

• Utilizing artificial neural networks (ANNs) trained on datasets of known scatterer shapes.
• Reconstructing object shapes from proximity electromagnetic field data.
• Comparing ANN-based methods with topology optimization and a hybrid ANN-topology optimization approach.

Main Results:

• ANNs can learn underlying scatterer structures from prepared datasets, aiding in non-unique inverse problem solutions.
• Topology optimization alone may not effectively recover scatterer geometry.
• A hybrid approach integrating ANNs and topology optimization demonstrates superior performance.

Conclusions:

• ANNs provide a powerful tool for incorporating prior assumptions into nano-optical inverse problem solutions.
• Hybrid methods combining ANNs with topology optimization offer the most promising strategy for accurate shape reconstruction.
• This research has significant implications for fields like optical metrology.