Chirped fiber Bragg gratings (CFBGs) are essential components in optical communications and ultrafast laser systems, providing critical functions such as chromatic dispersion compensation and pulse shaping. Achieving optimal performance requires precise control over two structural parameters, that is the local grating period distribution and the refractive index profile (apodization). While deep learning has recently emerged as a promising tool for inverse design, many existing approaches formulate the task as a parameter retrieval problem, mapping spectral data to a limited set of scalar coefficients based on predefined functions. Here, we present a flexible, data-driven inverse design framework using a multi-task U-Net architecture capable of reconstructing arbitrary, free-form apodization and chirp profiles. A key feature of our approach is the explicit utilization of both reflectivity and group delay spectra as inputs, which enhances the retrieval of phase information essential for accurate dispersion engineering. Furthermore, to ensure that the predicted structures remain smooth and physically realizable without imposing strict geometric constraints, we incorporate a composite loss function with Total Variation (TV) regularization. Numerical verification via the Transfer Matrix Method (TMM) demonstrates that the proposed model successfully reproduces complex target specification, offering a robust and versatile tool for advanced optical filter design.
Lee et al. (2026) studied this question.