AI-Driven Optimization of Multilayer Thin-Film Structures for Advanced Optical Applications

Designing multilayer thin-film structures with tailored optical responses is a complex and computationally intensive task, often requiring repeated simulations and limited physical intuition about the role of each layer. In this study, we propose a data-driven framework that leverages a fully connected neural network (FCNN) as a surrogate model to accelerate and interpret the design of a high-reflectivity multilayer structure composed of alternating Ta2O5 and MgF2 layers. The FCNN model is trained to predict reflectance spectra with high accuracy over a wide design space defined by the thicknesses of individual layers. Beyond fast prediction, we analyze the physical influence of each design parameter using Principal Component Analysis (PCA), Global Sensitivity Analysis, and SHapley Additive exPlanations (SHAP). These techniques provide deep insights into layer significance, symmetry, and redundancy in the design, shedding light on the complex relationships between structure and spectral response. Finally, we integrate the surrogate model with the Non-dominated Sorting Genetic Algorithm II (NSGA-II) multi-objective optimization algorithm to maximize both the transmission peak and its full width at half maximum (FWHM) for a desired wavelength band. We then elect the best solution using a Multi-Criteria Decision-Making (MCDM) approach. This work demonstrates the synergy between AI and photonic design, enabling efficient optimization and facilitating physical interpretability. The framework provides a robust pathway for the intelligent design of optical filters and mirrors, with potential applications across telecommunications, sensing, and laser systems.