A new neural network splits turbulence prediction into coarse and fine halves, and that split alone cuts training memory by nearly 80 percent.
The model, called ScaleSplit-NO, is built from two smaller networks instead of one giant one. A Parent network predicts a low-resolution version of a 3D fluid field one step into the future. A Child network then fills in full-resolution detail on local patches, conditioned on what the Parent guessed. Neither network ever has to process the entire full-resolution field at once, which is the actual trick behind the memory savings. The Child is pretrained on its own first, then bolted onto the Parent through connections that start at zero, so the combined system does not need retraining from scratch.
On a 256-cubed turbulence benchmark known as JHTDB256, the combined model's error came in 53% lower than the strongest baseline, while training memory dropped 79% compared to the most memory-efficient baseline tested. Full 3D turbulence simulation is normally done with brute-force numerical solvers that are accurate but slow and expensive to run at high resolution, which is also why training data for neural replacements is scarce in the first place. Decoupling resolution from memory cost means high-resolution surrogate models become trainable on hardware that could never fit a full-resolution model in memory, and a single pretrained Child can be reused across different Parents for extra accuracy without retraining.
The researchers also ran it on a real-world case: predicting wind flow through a grid covering an actual Montreal district, cutting one-step error by 65.8% versus baseline. That is a reminder this is aimed at practical jobs like urban wind engineering, not just leaderboard benchmarks.