Anomalous Diffusion in Photonics

A drop of ink spreads because every particle is being knocked around at random. Nothing steers it, yet the cloud grows at a completely predictable rate: the mean-square displacement rises linearly with time.

Follow a single particle and you get a Brownian trajectory: jagged at every scale, each step independent of the last. This is the textbook picture, and for a century it was enough.

Inside a living cell the same particle behaves differently. It gets stuck in crowded pockets for long stretches, then makes a sudden long flight. Same number of steps, entirely different statistics. This is anomalous diffusion.

Plot mean-square displacement against time on log-log axes and the exponent α becomes a slope you can read straight off. α = 1 is ordinary. Anything else is the cell telling you something. The catch: this needs long, clean trajectories, and nobody has those.

Real traces are a few dozen blurred points. Give them to a network trained on millions of simulated walks and it recovers α and the underlying model directly, well past where a slope gives up. We co-organised the AnDi Challenge to find out whose methods actually worked, twice, and published both answers in Nature Communications. The honest summary: the machines won, and we still cannot fully explain how they do it.