Two different ways to be wrong
An AI simulation can follow a system plausibly for a while, then produce values that grow without limit. In weather and fluid models, such artificial runaway behaviour can make a long simulation useless. A second problem remains even when that does not happen: the simulated path can still differ from reality.
MIT’s 20 November research account highlights a way to address the first problem. Sunbochen Tang, Themistoklis Sapsis and Navid Azizan developed a constraint that guarantees bounded trajectories for the mathematical model they construct. Bounded means the model’s states stay within finite limits; it does not mean every state is correct.
A rule built into the model
The system learns both how the dynamics evolve and an energy-like measure. An added component, called a projection layer, adjusts the predicted motion so that this measure decreases outside a learned boundary. The formal guarantee follows from that construction and its mathematical conditions, rather than from a penalty that merely encourages good behaviour during training.
The open author paper examines Lorenz systems and a reduced-order Kuramoto–Sivashinsky model, established mathematical tests of chaotic behaviour. It reports better preservation of long-term statistical patterns than an otherwise comparable unconstrained model. These experiments are not a test of tomorrow’s local weather forecast.
What the guarantee leaves open
Chaos makes exact long-range prediction difficult because tiny differences in starting conditions can produce very different paths. Researchers may instead ask whether a simulation captures the distribution of states over time. Preventing runaway values helps make that assessment possible; boundedness alone does not prove those statistics match reality.
The authors’ repository provides evaluation notebooks and checkpoints, which other researchers can inspect and test. Model Current has not run them. MIT says testing on realistic weather systems remains further work. When reading a claim of reliable scientific AI, ask which property is guaranteed, under what conditions, and which real-world application has actually been evaluated.
Sources
- MIT News: Physically constrained models of chaotic systems20 November research account explains hard versus soft constraints and identifies realistic weather models as further work.
- Tang, Sapsis and Azizan: Open author versionEarlier preprint, revised December 2025; full methods, boundedness theorem, Lorenz benchmark experiments, reduced-order Kuramoto–Sivashinsky model and limitations.
- Azizan Lab: Dissipative chaos research repositoryAuthors’ public code, setup instructions, evaluation notebooks and checkpoints. Availability was inspected; experiments were not rerun.



