通用长期物理仿真的潜在生成求解器

This research presents a method for long-horizon surrogate simulation across heterogeneous PDE systems, termed Latent Generative Solvers (LGS). The approach employs a two-stage framework: the first stage maps diverse PDE states into a shared latent physics space using a pretrained Variational Autoencoder (VAE), while the second stage learns probabilistic latent dynamics through a Transformer trained via flow matching. A key mechanism is an uncertainty knob that perturbs latent inputs during training and inference, enabling the solver to correct off-manifold rollout drift and stabilize autoregressive predictions. Additionally, flow forcing is utilized to update a system descriptor (context) from model-generated trajectories, aligning training/testing conditions and enhancing long-term stability. Pretraining is conducted on a curated corpus of approximately 2.5 million trajectories.