Inverse Design of Thin-Film Optical Characteristics with Guided Flow Matching

Sammanfattning

Thin-film materials are fundamental components in modern technology and are widely used in applications such as solar cells and other optoelectronic devices. Their optical response, which depend on both the dielectric functions and thickness of the constituent materials, can be measured by hyperspectral imaging (HSI) cameras. Since the complex dielectric functions of thin-film materials determine many properties relevant to optoelectronic applications, accurately inferring them from optical measurements is of significant interest. This motivates both accurate modelling of the forward process from material properties to optical measurements and solving the corresponding inverse problem of recovering material properties from measured optical responses. In this work, we construct a physics-based simulation pipeline based on the Tauc-Lorentz oscillator model for dielectric functions, the Bruggeman effective medium approximation, and the transfer matrix method to generate synthetic data consisting of dielectric functions and corresponding reflectance-transmittance (RT) images. Using this dataset, two classifier-free guided flow models were implemented and trained with the conditional flow matching objective and a Gaussian conditional optimal transport probability path. The first model emulates the physical forward process from dielectric functions to RT-images, while the second performs the inverse mapping from RT-images to dielectric functions. Both models employ U-Net-style architectures to parametrise the learned vector fields. Experiments show that the forward model can emulate the physical simulation pipeline for materials well represented in the training data, achieving mean absolute differences at or below the the approximate one-percentage-point signal-to-noise ratio of an HSI camera. The inverse model is able to recover in-distribution dielectric functions with high precision, and RT-images simulated from generated dielectric functions remain largely indistinguishable from those obtained from the true values. While the present work considers a simplified setting based on three spectral bands, it provides a scalable foundation for extension to more general configurations. Future work may therefore extend the method to the full HSI spectral range and further fine-tune it on real measurement data.