VAEs Uncover Latent Structures in Large-Scale X-ray Scattering Data.

Scientific facilities are generating X-ray scattering data at rates that overwhelm traditional processing methods. To address this, a new approach utilizes a domain-specific attention-based Convolutional Variational Autoencoder (C-VAE) trained on 1.5 million X-ray scattering images. This model learns compact, low-dimensional representations that effectively capture structural variations under diverse experimental conditions. The learned latent space from the C-VAE organizes data into well-defined clusters and smooth trajectories, which directly reflect the progression of experiments. This capability also extends to generating synthetic scattering images that represent various structural states. The model has been successfully deployed without retraining to analyze time-resolved film formation experiments at two synchrotron facilities, revealing interpretable latent structures. Benchmarking against DINOv3, a general-purpose vision foundation model, demonstrated that the domain-specific training of the C-VAE yields a more interpretable latent organization for X-ray scattering data. Both offline exploration and live analysis workflows are integrated into the Latent Space Explorer, part of the MLExchange platform, providing interactive tools for structural investigation.