Physics-constrained neural networks for surrogate modeling of lossless periodic structures