<p>The field of photonics has experienced significant progress due to the exploration and integration of machine learning techniques that can be used to model complex electromagnetic behavior as a function of wavelengths and design parameters. Data samples are extracted from electromagnetic simulators, and the resulting data are used by machine-learning algorithms to build efficient and accurate surrogate models that can be used to speed up design tasks. In this work, we propose a novel hybrid technique that integrates supervised and unsupervised learning. The wavelength variable is included in the input space of the design parameters, and a clustering approach identifies a set of wavelength intervals. This interval information is assigned to a categorical variable and encoded as an input. Finally, a supervised deep learning model is built to link the input space with the electromagnetic behavior of interest. We also explore how including physics-informed regularization smooths the model’s predictions. Extensive numerical experiments validate the method.</p>