Kastor: An efficient fine-tuning strategy for generative emulation of PDE simulations

What the math gives to ML

The paper contributes a nonstandard constraint for conditional generative emulators: the latent noise mean should map to the conditional mean of the next physical state. This anchors an otherwise stochastic generator at a predictable, low-variance output and can reduce rollout drift without eliminating stochastic samples away from the anchor. The construction transfers to diffusion, flow, GAN, and latent-variable sequence models whenever the model accepts explicit noise conditioning. The most direct experiment is to add the mean-prediction regularizer to a conditional generative forecaster and measure long-horizon error, calibration, and sample diversity against the unregularized objective.