Designing a multilayer dielectric thin film stack to hit a target optical response is normally a slow trial-and-error process, or a computationally expensive sweep with gradient-based optimizers. In this project, I implemented Res-GLONet, a residual generative optimization network, to inverse-design 1D thin film stacks directly from a target reflection or transmission spectrum. The generator, a residual multilayer perceptron network with 16 cascaded residual blocks, takes a random noise vector and outputs the layer thicknesses and a probability distribution over a material library, from which the refractive index of each layer is sampled. This design is fed to a transfer matrix method (TMM) solver, which acts as the discriminator by computing the resulting spectrum and the loss against the target response.

Since the reference TMM implementation was written in NumPy, I reimplemented and vectorized it in PyTorch using scattering matrices, so the solver could sit inside the network’s computation graph and backpropagate gradients directly, batched across devices, wavelengths and incidence angles. I also found and corrected a sign error in the published loss function that would otherwise cause training to diverge. The model was tested on two designs: a 3-layer anti-reflection coating, which converged to an average reflectivity of about 2.1% (within 1% of the benchmark paper’s results), and a 50-100 layer narrowband transmission filter targeting sharp transmission peaks at 850 nm and 870 nm, drawn from a 7-material library, which converged to about 10% mean squared error. Both designs converged within about 10 minutes on a personal laptop with a modest GPU, with PCA plots of the design space showing the generated samples collapsing from a broad random spread down to a tight cluster near the optimum over the course of training.