Bidirectional autoregressive latent diffusion predicts MHD fields with self-supervision

Alexander Scheinker presents a novel bidirectional autoregressive latent diffusion approach for solving forward and inverse problems in magnetohydrodynamics (MHD). The model simultaneously predicts multiple coupled fields—mass density, pressure, velocity, and magnetic field components—by learning their temporal evolution in a latent space. A key innovation is the self-supervised consistency metric: by flowing forward in time from an initial state and then reversing the process, the model compares the predicted fields at each step. The degree of agreement between forward and backward predictions serves as a proxy for uncertainty and error, eliminating the need for ground truth data during inference. This makes the model particularly valuable for scenarios where direct measurements are unavailable or expensive.

The method also demonstrates potential as a non-invasive plasma diagnostic tool. By leveraging adaptive feedback from sparse diagnostics or limited sensor views, the model becomes more robust and can reconstruct full-field dynamics from partial observations. This is critical for applications like fusion reactor control, where internal plasma parameters are difficult to measure directly but are essential for stability. The work bridges generative AI and physics-based modeling, offering a practical framework for uncertainty-aware predictions in complex dynamical systems. Future extensions could include real-time plasma monitoring in tokamaks or space weather forecasting, where accurate, uncertainty-quantified predictions are paramount.