Disentangled Compositional Diffusion for Controllable Scientific Data Generation

Diffusion and flow-based generative models are capable of generating high quality samples, but their internal representation is difficult to manipulate for controllable and interpretable generation, particularly for scientific applications. We employ decomposed diffusion models, which encode data into a latent representation, enabling interpretability by disentangling individual components and controllability by selecting components with desired features to generate new samples. We demonstrate applications on different datasets, including interpretability and controlled generation on temperature fields and vortex-induced flow fields, synthesizing human aortic geometries from very few samples, and 2D turbulent flow fields.