Deep learning outperforms RCWA for multimodal measurement of high-aspect-ratio nanostructures

Measuring performance-critical parameters in nanostructures is essential for optimizing optics. Our previous work replaced rigorous coupled-wave analysis (RCWA) with a multi-layer perceptron (MLP) that replicated RCWA Mueller matrix spectroscopic ellipsometry (MMSE), enabling gradient-based fitting at inference. While this measured tilt well, it was restricted to a 2D geometry and validated only for critical angle transmission (CAT) bar tilt. We later extended it to a 3D model for multiple parameters, achieving an R² of 0.88 compared ground truth. Still, the MLP approach required separate models for each parameter, and each inference took about 5 seconds. We propose a self-supervised pipeline that directly maps MMSE spectra to parameters, combining a variational autoencoder (VAE) and a convolutional neural network (CNN) with skip-connections. We pre-train the VAE on experimental MMSE spectra to learn a latent space. We then feed the encoder output and original spectra into the CNN, using SAXS and edge-detection outputs as labels, self-supervised with synthetic fits. Our approach achieves an R² of 0.94 for both tilt and profile prediction and reduces inference to under one second.