A recurring problem in AI models of chaotic physical systems is that small prediction errors can compound until the model produces values that no real system could reach. MIT News reported that an MIT-led research team has proposed a way to prevent that failure mode by building hard physical limits directly into the model.

According to MIT News, the work comes from Sunbochen Tang, Navid Azizan and Themistoklis Sapsis, affiliated with MIT and its Laboratory for Information and Decision Systems. MIT News said the findings were published in the Proceedings of the National Academy of Sciences.

The report frames the advance around boundedness: even when a system is chaotic, its behavior still stays within physical limits. The researchers’ method, as described by MIT News, aims to preserve the complicated dynamics of chaos while preventing the model from running away into impossible states during long simulations.

MIT News said the method differs from earlier approaches that mainly used training penalties or incentives intended to encourage physically plausible behavior. In the MIT team’s approach, the model learns an energy-like function and a boundary that confines the chaotic behavior. A projection layer then adjusts the model’s predicted evolution so that, outside that boundary, the energy-like quantity must decrease. MIT News reported that this setup is meant to provide a theoretical guarantee that predictions remain bounded over arbitrarily long forecast windows.

That distinction matters because a model can fit short-term data yet still fail badly when run forward for longer periods. MIT News said the researchers argue that simply capping outputs at arbitrary values would keep predictions bounded, but could also prevent the model from representing important structure that should be learned from data.

According to MIT News, the team tested the approach on several well-known physical systems, from simpler textbook examples to more complex multidimensional cases. MIT News reported that the constrained models stayed close to the true outcomes, while conventional methods eventually diverged into impossible predictions.

The reported applications span systems in which chaotic dynamics complicate forecasting and control, including weather and climate patterns, turbulent flow, aeronautics, hydrology and industrial processes. MIT News said the projection layer is modular and model-agnostic, meaning the researchers expect it to work with different neural-network architectures.

MIT News also said the researchers see particular promise where data are limited, because physical constraints can supply additional information beyond the training set. As a next step, the team plans to test the method on larger systems, including data-driven models of realistic weather systems, according to the report.

MIT News reported that the paper and open-source code are accessible online. It also said a follow-up work by members of the team was presented this year at the Learning for Dynamics and Control conference, where it received the conference’s Best Paper Award.