Many scientific problems require inferring hidden properties from measurements, like reconstructing an object from waves. Traditional solvers are computationally expensive because they repeatedly solve complex physics equations.
AdjointFlow creates a learned map from measurements to reconstructions. It starts from a physics-based approximation rather than random noise, and uses a specially designed transport that remains stable even when the target distribution is very concentrated.
On several scientific tasks, AdjointFlow produced more accurate reconstructions than conditional diffusion models, while being up to five orders of magnitude faster than plug-and-play methods that iteratively query the physics model. This makes high-quality inversion practical.