Neuroevolution achieves optimized chiral metasurfaces to advance nanophotonics design

Designing chiral metasurfaces with precise optical properties is a major hurdle in nanophotonics. This is because their shapes and responses are complexly related, making traditional design methods slow and inefficient. Davide Filippozzi, Arash Rahimi-Iman and colleagues at Justus-Liebig-Universität Gießen are tackling this challenge by integrating a powerful evolutionary algorithm known as NEAT into their machine learning framework. This innovative approach allows the system to independently develop optimal neural network architectures, thereby avoiding the need for manual design and increasing the speed and accuracy of metasurface optimization. The research team demonstrated that these automatically evolved networks not only match the performance of traditionally designed models, but also offer higher efficiency and the potential for transfer learning between simulations and real-world experiments, paving the way for adaptive, self-configuring systems for automated photonic design.

NEAT optimizes chiral metasurface inverse design

Scientists have developed a machine learning framework for designing chiral metasurfaces, light-manipulating structures, by integrating the NeuroEvolution of Augmenting Topologies (NEAT) algorithm with a deep learning optimization pipeline. This work addresses the challenge of efficiently designing these complex structures, where the relationship between shape and optical properties is nonlinear and traditional methods often require extensive manual tuning. The team designed a system in which NEAT autonomously evolves both the neural network's topology and connection weights, creating task-specific architectures without human intervention, while reinforcement learning strategies refine the model's understanding of the design space. This study pioneered the use of a dataset of 9,600 simulated gallium phosphide metasurface shapes generated by computational modeling to evaluate the performance of NEAT under various conditions.

The researchers systematically varied the input dimensions, feature scaling method, and data size to determine the optimal configuration for the evolving neural network and found that standardized feature scaling consistently yielded the best results. As a result, the neural networks evolved by NEAT, characterized by their compact size, demonstrated predictive accuracy and generalization comparable to, and in some cases exceeding, the dense networks used initially. In our experiments, we used rigorous coupled-wave analysis simulations to model the spectral properties of the metasurface and provided the basis for machine learning algorithms to learn the complex relationships between structure and optical response. This approach successfully predicted metasurface designs exhibiting strong circular dichroism in the visible spectrum and, importantly, enabled transfer learning between simulated and experimentally fabricated structures, demonstrating potential for real-world applications. This work enables a scalable path towards an adaptive self-configuring machine learning framework for automated photonic design and provides a powerful tool for rapidly prototyping and creating metasurfaces with customized optical features.

Automated chiral metasurface design with NeuroEvolution

Scientists have achieved a breakthrough in the automatic design of chiral metasurfaces, nanoscale structures that interact with light, by integrating advanced machine learning techniques called NeuroEvolution of Augmenting Topologies (NEAT) into existing deep learning optimization frameworks. This work overcomes a central challenge in nanophotonics, namely the complex relationship between the shape of a metasurface and its optical properties, by enabling automatic configuration of neural network architectures tailored to specific design tasks. The team approach eliminates the need for manual network design and allows the model to dynamically adapt its complexity to the optimization problem. The experiment included a dataset of 9,600 simulated gallium phosphide metasurface geometries, each carefully designed and analyzed using electromagnetic simulation to determine the optical response.

The researchers tested NEAT under different conditions, such as different input dimensions and data sizes, and investigated feature scaling methods, especially regularization and standardization, to optimize training efficiency and convergence. Results show that standardized feature scaling consistently yields the best performance, predicting both strong circular dichroism and preferred handedness reflectance. Despite being relatively compact, the neural networks evolved by NEAT achieved predictive accuracy and generalization comparable to, and in some cases exceeding, manually designed networks. Measurements confirm that these models successfully predict metasurfaces exhibiting strong circular dichroism in the visible spectrum and can transfer learning from simulated data to experimental data. This capability paves the way for adaptive, self-configuring machine learning frameworks that can accelerate the design of photonic devices and potentially couple data-driven design with automated manufacturing processes.

NEAT designs high-performance chiral metasurfaces

This study demonstrates a new approach to designing chiral metasurfaces, nanoscale structures that interact with light, by integrating a technique called NeuroEvolution of Augmenting Topologies (NEAT) into a machine learning pipeline. The team was able to use NEAT to automatically design the architecture of a neural network, eliminating the need for manual tuning typically required by these systems. These evolved networks accurately predicted the optical properties of metasurfaces and achieved comparable or improved performance compared to networks with manually designed structures. The resulting model is not only accurate but also resource efficient, successfully predicting metasurfaces exhibiting strong circular dichroism, an important property for many optical applications, and enabling transfer of learning from simulated designs to real-world experimental data.

Our research revealed that standardized feature scaling consistently improves performance, suggesting that careful data preparation improves the efficiency of the neuroevolution process. The researchers highlight the benefits of three aspects of their approach, including data evolution, structural design evolution, and machine learning model architecture evolution, all working together to optimize the design process. The authors acknowledge that the performance of the NEAT algorithm is affected by factors such as input dimensionality and data size, so careful consideration is required when applying this technique to different problems.