MIT AI Keeps Chaotic Models Within Physical Bounds

Researchers at MIT have found a way to stop artificial intelligence models of chaotic systems drifting into predictions that could never happen in the real world, in work the university says offers the first formal guarantee of its kind.
The team, from MIT's Laboratory for Information and Decision Systems (LIDS), announced the approach on Friday 20 November. Their paper, by graduate student Sunbochen Tang and professors Navid Azizan and Themistoklis Sapsis, was published in the Proceedings of the National Academy of Sciences.
Why chaotic models run away
Chaotic systems — weather and climate patterns, turbulent flows in aviation and ocean currents — are exquisitely sensitive to their starting conditions. As Azizan put it: "Tiny differences in the starting conditions can lead to dramatically different outcomes in a chaotic system."
That sensitivity makes long-range prediction hard, and it creates a subtler problem for machine-learning models. A model can match the data over a short period and still fail when it is left running, as small errors accumulate until the simulation produces values that grow without limit — predictions that could never occur in a real system. "Chaos does not mean anything can happen," Azizan said. "These systems often remain bounded, and their models should, too."
Energy-based limits, built in
Earlier attempts to keep models realistic usually worked by imposing penalties or incentives during training. Tang explained that those only apply while the model is being trained, so "there's no guarantee that the model will satisfy the underlying constraints intended by such incentives at test time" once it faces new inputs.
The MIT approach is different. The model learns an energy-like function, based on how energy dissipation keeps real chaotic systems bounded, along with a boundary where its chaotic behaviour should be confined. A projection layer then adjusts the model's predictions to enforce energy dissipation as a hard constraint: outside the boundary, the energy-like quantity must decrease, drawing the simulated system back toward the bounded region.
Simply capping predictions would not work, Tang said, because an arbitrary limit — "for example, by setting an arbitrary limit on physical quantities, such as 100 degrees for temperature" — could stop the model learning complex features that should come from the data. The new method, he said, "gives the model room to learn complex dynamics while guaranteeing that its trajectories stay bounded".
Tested on textbook and complex systems
The team tested the framework on a range of well-known physical systems, from simple textbook cases to highly complex multidimensional ones, and found it closely matched the true outcomes while conventional approaches soon deviated into impossible predictions. The tests included the Lorenz 96 system and a reduced-order model of the Kuramoto-Sivashinsky equation, both standard benchmarks for chaotic dynamics.
The projection layer is modular and model-agnostic, so it can work with different neural-network architectures, and the team has made the paper and the open-source code freely available. A follow-up work by members of the team, presented this year at the Learning for Dynamics and Control conference, won the conference's Best Paper Award.
The researchers now plan to test the approach on ever-larger systems, including data-driven models of realistic weather systems. "Trustworthy and reliable AI needs more than just evidence that a model usually works," Azizan said. "For high-stakes applications, you need guarantees about the properties that must always hold. In this work, we established one such guarantee: boundedness."