<ul><li>Autoregressive next-step prediction models are widely used for building data-driven neural solvers for time-dependent partial differential equations (PDEs).</li><li>A new approach proposes a latent diffusion model for PDE simulation, reducing computational costs.</li><li>Using an autoencoder, different types of meshes are mapped onto a unified structured latent grid, enabling the capture of complex geometries.</li><li>The proposed model outperforms deterministic baselines in accuracy and long-term stability, demonstrating the potential of diffusion-based approaches for robust data-driven PDE learning.</li></ul>

Generative Latent Neural PDE Solver using Flow Matching

Discover more