Physics-Informed Machine Learning for Predicting SAXS Data of Lyotropic Liquid Crystals using Generative Models

Machine learning effectively models complex experimental data for soft materials, but limited availability of data can hinder training. To address this challenge, generative models are combined with physics-informed machine learning to predict Small-Angle X-ray Scattering (SAXS) data of lyotropic liquid crystals composed of a solution of Pluronic® F127, NaF, and water. We employ a variational autoencoder (VAE) to generate synthetic SAXS training data and develop a neural network with a physics-informed layer that integrates the Debye function, Bessel function, and spherical harmonics. Trained on only six experimental samples, our model predicts SAXS intensity curves for both an unseen sample within and beyond the original data range.