Supervisors

Suitable for

Abstract

Prerequisites: AI/ML, reinforcement learning, Python

 

Background

• Photonic materials and devices, such as metasurfaces and photonic crystals, are crucial in applications ranging from telecommunications to biomedical sensing. Traditional approaches to designing these structures often rely on extensive trial-and-error simulations, which can be time-consuming and computationally expensive. By using ML and AI, it becomes possible to explore large design spaces more efficiently, leading to novel photonic structures with optimized performance.
• AI-based photonic design typically involves:

1. Surrogate modelling: Training a model (e.g., a neural network) to predict optical responses (reflection, transmission, bandgap properties) from geometrical parameters.

1. Inverse design: Learning to propose geometry configurations that meet specific optical targets directly.

• Recent work in this field has shown significant reductions in computational cost and enhancements in device performance, making AI-driven photonic design an increasingly promising research direction.
• This project will develop and evaluate a machine learning approach for the inverse design of a particular class of photonic structures (e.g., photonic crystals, metasurfaces). The main question is how effectively AI-based methods can discover device geometries that meet predefined optical goals compared to established optimization methods.

 

Focus

• Expected Contributions and goals include:

1. ML Model Training:

• - Implement a forward model (predicting optical response from geometry) and/or an inverse model (predicting geometry from target responses).
• - Assess prediction accuracy using relevant metrics.

1. Baseline Comparison:

• - Compare AI-driven design strategies with a conventional optimization approach (e.g., genetic algorithm or gradient-based method).
• - Evaluate time-to-solution, final device performance, and robustness.

1. Insights and Design Principles:

• - Derive insights into how AI-driven solutions differ from conventional methods.
• - Highlight potential new strategies or patterns in photonic device engineering.

 

Method

We expect to explore reinforcement learning or Bayesian optimization to iteratively refine designs with fewer required simulations. Familiarity in these areas and AI/ML foundations is therefore required. Proficiency in implementing ML algorithms is also preferred.

 

References

[1] Chen, Mu Ku, et al. "Artificial intelligence in meta-optics." Chemical Reviews 122.19 (2022): 15356-15413.

[2] Bonfanti, Silvia, et al. "Computational design of mechanical metamaterials." Nature Computational Science 4.8 (2024): 574-583.

[3] So, Sunae, Trevon Badloe, Jaebum Noh, Jorge Bravo-Abad, and Junsuk Rho. "Deep learning enabled inverse design in nanophotonics." Nanophotonics 9, no. 5 (2020): 1041-1057.