What happened
A team at the Laboratory for Information and Decision Systems (LIDS) at MIT has developed a way to keep machine learning models of chaotic systems, including weather, climate and turbulent flows, from producing physically impossible forecasts, MIT News reported. The paper, which appears in the journal Proceedings of the National Academy of Sciences, is by graduate student Sunbochen Tang and professors Navid Azizan and Themistoklis Sapsis.
In the method, the model learns an energy-like function together with a boundary meant to contain its chaotic behaviour. An added projection layer enforces energy dissipation as a hard constraint, so beyond that boundary the energy-like quantity has to fall and the simulation is pulled back inside. The team says this provides a theoretical guarantee that predictions stay bounded over forecast horizons of any length.
In tests on well-known physical systems, from simple textbook examples to complex multidimensional ones, the approach tracked true outcomes closely, while conventional methods soon diverged. Open-source code is available alongside the paper. Azizan said the layer is model-agnostic. Tang said the team next plans to try it on data-driven models of realistic weather. A follow-up paper by team members earned the Best Paper Award at this year's Learning for Dynamics and Control conference.
Context
Chaotic systems are highly sensitive to their starting conditions, and small errors in learned models can compound until outputs grow without limit. Tang said earlier fixes relied on penalties or incentives during training, which offer no assurance once a model meets new inputs. He added that simply capping values at arbitrary limits can stop a model from learning complex features from the data.
Das Briefing
Die Tech-Woche in einer E-Mail.
Die lesenswerten Beiträge der Woche. Jederzeit abbestellbar.
