Problem

Designing chiral metasurfaces with specific optical properties, such as circular dichroism, is a significant challenge in nanophotonics. This is due to the highly nonlinear and complex relationship between the metasurface's geometric parameters and its chiroptical response. While machine learning pipelines have accelerated this design process, their effectiveness is often limited by the manual selection and tuning of neural network architectures, which can introduce bias and hinder adaptability.

Method

This research integrates the NeuroEvolution of Augmenting Topologies (NEAT) algorithm into an existing deep-learning optimization framework designed for dielectric chiral metasurfaces. NEAT automates the evolution of both the neural network's topology (structure) and connection weights, eliminating the need for manual architecture design. The framework uses a reinforcement learning strategy to iteratively improve the understanding of the solution space and fine-tune the evolved models. A dataset of 9,600 simulated Gallium Phosphide (GaP) metasurface geometries, generated using rigorous coupled-wave analysis (RCWA), was used to train and evaluate the NEAT-evolved networks. The study explored various configurations, including different dataset sizes (10%, 50%, 90%), input dimensionalities (6 coordinates for independent corners vs. 24 coordinates for all symmetry-related corners), and feature scaling methods (normalization vs. standardization). NEAT's fitness function was defined as the mean squared error (MSE) between predicted and simulated optical quantities, and experiments were repeated with different random seeds for statistical validity.

Results

The most accurate NEAT models consistently emerged from configurations using six input features (representing three independent corner coordinates) combined with standardized feature scaling. This setup yielded the lowest mean squared errors (MSEs) on unseen data, demonstrating improved generalization compared to networks trained with all 24 symmetry-related inputs. Standardization consistently outperformed normalization, reducing the average MSE by approximately 10% in the 6-input case. Furthermore, standardization led to the evolution of more complex network topologies with nearly twice as many active connections, suggesting it promotes richer connectivity patterns that capture nonlinear dependencies more effectively. NEAT also acted as an implicit feature selector, emphasizing the most informative geometric descriptors and suppressing redundant symmetric parameters in the 24-input configurations. When integrated into the full optimization pipeline, the best-performing NEAT-evolved networks achieved circular dichroism (ΔRCD) values up to 0.0095 and preferred handedness reflectance (Rpref) of approximately 0.016 for two-output configurations, and ΔRCD ≈ 0.0094 for single-output configurations. These results were comparable to or better than those obtained using manually defined neural network architectures.

Implications

This work demonstrates a scalable and adaptive approach to automated photonic design, allowing for the autonomous evolution of neural network architectures tailored to specific optimization tasks. By eliminating manual network engineering, the framework enhances predictive accuracy, generalization, and computational efficiency in designing chiral metasurfaces. The findings highlight the importance of standardized feature scaling in neuroevolution for complex physical problems. The developed pipeline has the potential for transfer learning with experimental data, linking design directly with fabrication. Looking forward, this approach can serve as a building block for agentic artificial intelligence (AI) frameworks, enabling fully autonomous and self-configuring design and production pipelines in nanophotonics and other fields requiring complex material optimization. The ability for machine-learning models to dynamically self-adjust to underlying physics could critically enable future data-driven workflows within autonomous labs.