A paper published 8 October in Nature Communications derives a diffusion model’s noise schedule from a known particle-spreading law. In a turbulent-plasma case study, the authors report that the surrogate reproduced the laboratory-measured variance scale and captured the excess-kurtosis trend in test-particle simulations.
Most generative diffusion models use a chosen schedule to decide how much Gaussian noise to add at each step. Reichherzer and colleagues instead start with a macroscopic variance law—the mean-square spread of a particle population over time—and set the schedule’s squared noise amplitude to the slope of that curve. That changes one scalar schedule in a standard diffusion model and aligns the model’s time coordinate with physical spreading.
The schedule fixes the variance, so the neural network can learn the remaining distributional shape, including tail structure inherited from the initial distribution. Training uses samples from that initial particle distribution and synthetic Gaussian corruptions; it does not require physical transport observations at intermediate times.
For charged particles moving through a turbulent magnetic field, the prescribed variance law captures a transition from early ballistic spreading, where variance grows roughly with time squared, toward later diffusion, where it grows roughly linearly. The paper compares the resulting variance scale with proton-radiography measurements from a laboratory experiment. Separately, the surrogate excess-kurtosis curve—a measure of distributional tail weight—tracks the component-kurtosis trend in test-particle simulations across 20 independent runs.
The model represents particle distributions, not individual paths. It does not enforce the particles’ finite speed of travel or preserve how one particle’s motion at one moment relates to its motion later. Its Gaussian approximation works best when many small, nearly independent deflections make the ensemble’s overall spread roughly bell-shaped.
Sources
- Patrick Reichherzer et al., “Generative diffusion surrogates with analytical variance schedule,” Nature Communications 17, 10408 (2026), published 8 October 2026.
- Full paper, arXiv:2609.01705v1, especially §§1–2.5 and Fig. 4.
- University of Oxford announcement, 8 October 2026.
- Figure 1 from Reichherzer et al., CC BY 4.0.