AI models used to predict chaotic systems like weather and fluid turbulence actually fail if they are too 'accurate,' proving that randomness is a physical requirement for realistic long-term forecasting.
March 31, 2026
Original Paper
Stochasticity and probabilistic trajectory scoring are essential for data-driven closures of chaotic systems
arXiv · 2603.28671
The Takeaway
This discovery challenges the assumption that higher precision always leads to better science. It proves that in chaotic systems, being perfectly 'deterministic' suppresses natural variability, meaning that injecting randomness is actually necessary to keep the model's physics from breaking.
From the abstract
Coarse-grained models of chaotic systems neglect unresolved degrees of freedom, inducing structured model error that limits predictability and distorts long-term statistics. Typical data-driven closures are trained to minimize error over a single time step, implicitly assuming Markovian dynamics and often failing to capture long-term behavior. Recent approaches instead optimize losses over finite trajectories. However, when such trajectory-based training is carried out with deterministic pointwi