Researchers at MIT say they have found a way to keep machine-learning models of chaotic systems from drifting into predictions that could never happen in the real world. According to MIT News, the method builds physical limits directly into the model, so its forecasts stay within realistic bounds no matter how far into the future it runs.
The work, published in the Proceedings of the National Academy of Sciences, comes from MIT's Laboratory for Information and Decision Systems. The team is making the information behind the approach freely available.
Weather, climate, turbulent airflow around aircraft and ocean currents are all chaotic: tiny differences in starting conditions can lead to wildly different outcomes. That makes long-range prediction hard. Machine-learning models of such systems have improved a great deal in recent years, but MIT News reports they still tend to wander into physically impossible results over time.
Navid Azizan, a professor at MIT, explained the problem this way: a model can match observed data over a short stretch and still fail when it is left to run for a long time. Small errors pile up until the simulation produces values that grow without limit.
Azizan also stressed that chaos has limits. In his words, these systems often remain bounded, and their models should too.
Earlier attempts to prevent unrealistic output often relied on penalties or incentives applied while the model was being trained. Sunbochen Tang, an MIT graduate student and the paper's lead author, said those only shape training. Once the model meets new inputs in actual use, nothing guarantees it will respect the intended constraints.
Simply capping predictions at a fixed value, such as 100 degrees for temperature, would keep numbers bounded, Tang said, but it could stop the model from learning subtle and complex features straight from the data.
The MIT team took a different route based on energy. The model learns an energy-like function together with a boundary inside which its chaotic behavior should stay. The researchers then add a projection layer, a component that adjusts how the model predicts the system will evolve, so that energy dissipation becomes a hard requirement. Outside the boundary, the energy-like quantity must fall, which pulls the simulated system back toward the allowed region.
Because of that design, Tang said, the team can show theoretically that predictions remain bounded over arbitrarily long forecast windows. Azizan described the difference as hard constraining versus soft constraining: many existing methods encourage a model to follow the rules, while this one is built on hard-constrained neural networks. MIT News says this is the first work to establish such formal guarantees for machine-learning models of chaos.
The researchers tried the method on a range of well-known physical systems, from simple textbook examples to highly complex multidimensional ones. According to MIT News, the new approach closely matched the true outcomes, while conventional approaches soon veered into impossible predictions. The tests were on established benchmark systems, not on live weather or climate forecasts.
Reliable models of chaotic flows matter in many places. MIT News notes that turbulence can disrupt air travel, reduce the efficiency of wind turbines and affect vehicle fuel economy, and that the chaotic atmosphere makes long-term weather forecasting difficult. The same kinds of models are also used in industrial processes, aeronautics and hydrology.
The researchers argue that guaranteed bounds could make these tools more dependable. Themistoklis Sapsis, a professor of mechanical engineering, said the lack of such guarantees has long been a major obstacle to data-driven models of complex nonlinear systems. He said the approach opens the door to using the most advanced machine-learning architectures for prediction and control, even in systems that are inherently unstable.
The result does not make forecasts of chaotic systems perfectly accurate. It addresses one specific failure, outputs that run away to impossible values, and leaves the underlying difficulty of predicting chaos in place.
Around Prague, based on reporting by MIT News