Nanophotonic Design

Photonic devices are governed by a relationship between nanoscale geometry and optical function. Small changes to a structure’s shape, feature size, material distribution, or arrangement can strongly affect how light is guided, scattered, focused, filtered, or converted. As a result, conventional intuition-driven design can become increasingly difficult for compact, multifunctional, or highly constrained devices.

Our computational design research develops computational frameworks that allow optical functionality and not a predefined geometry to guide device discovery. By combining physics-based electromagnetic modelling with topology optimisation, shape optimisation, inverse design, and machine learning, we explore large and often non-intuitive design spaces to identify high-performance photonic architectures.

The performance of a photonic device is determined by its nanoscale geometry and the way these elements interact with propagating light. Even small geometric changes can produce large changes in optical response so designing compact or multifunctional devices often requires exploring a vast and highly non-intuitive space of possible structures. Our computational design research addresses this challenge by allowing the desired optical functionality and not a predefined device geometry to guide the search for an optimal design.

Starting from a target response we use inverse-design methods to identify structures that realise the required input–output optical mapping. Topology and shape optimisation play a central role in this process, treating the arrangement of material within a design domain as a set of continuously variable degrees of freedom at the device and unit cell level. Structures can therefore evolve toward configurations that may differ substantially from conventional component design.

Machine learning further expands the scope and efficiency of this design process. Deep neural networks can learn the complex relationship between geometry and optical response, providing fast surrogate models for performance prediction and enabling generative approaches in which suitable structures are proposed directly from a target specification. Reinforcement learning offers a complementary strategy. An agent interacts with a simulated photonic environment and progressively improves its design decisions through feedback. This is particularly valuable for complex, multi-parameter, and multi-objective problems, where design choices are strongly coupled and the optimal solution may be difficult to reach through manual iteration alone.

Across these approaches, our goal is to develop design frameworks that are not only computationally powerful but also practical. By incorporating fabrication constraints, tolerance to manufacturing variation, and material-platform considerations into the design process, we aim to create metasurfaces and integrated photonic components whose optimised geometries translate reliably from simulation into experimental devices. These methods provide a foundation for compact, robust, and multifunctional photonic systems for quantum technologies, sensing, and optical information processing.