MIT researchers have published a method that guarantees AI models of chaotic systems never drift into physically impossible states, no matter how long the simulation runs. The work, led by graduate student Sunbochen Tang and professors Navid Azizan and Themistoklis Sapsis at the Laboratory for Information and Decision Systems, appears in the Proceedings of the National Academy of Sciences [1][2]. The core advance is a projection layer that enforces energy dissipation as a hard constraint inside the model architecture, rather than relying on soft penalties during training [1][5].## Why current models fail on long horizonsChaotic systems such as weather, climate, and turbulent flows are highly sensitive to initial conditions, so tiny differences produce dramatically different outcomes [1][6]. Neural networks can match short-term data well, but small errors accumulate until simulated values grow without limit, even though the real system remains bounded [1][5]. Azizan notes that a model can look accurate for a short period and still fail when run longer, producing predictions that could never occur in reality [1][6]. This drift invalidates long-term statistics, which are often the primary goal in climate and turbulence modeling [2].## Hard constraints versus soft penaltiesPrevious attempts used penalties or incentives during training to encourage physical consistency [1][5]. Tang explains that these soft constraints only apply while the model is being trained; once deployed on new inputs, there is no guarantee the model will respect the intended limits [1][5]. Simply capping outputs at arbitrary values, such as forbidding temperature above 100 degrees, also fails because it can prevent the model from learning complex features present in the data [1][6]. The new approach embeds the constraint structurally so that every prediction, even on unseen cases, satisfies the physical bound [1][8].## How the projection layer worksThe framework simultaneously learns a dynamics emulator and an energy-like representation of the system [2][5]. A modular projection layer then adjusts the predicted evolution: outside a learned boundary, the energy-like quantity must decrease, pulling trajectories back toward the bounded region while allowing chaotic variation inside it [1][2]. This construction enforces dissipation by design and provides formal trajectory-boundedness guarantees over arbitrarily long forecast windows [2][5]. The layer is model-agnostic, so it can be added to different neural-network architectures without redesign [2][5].## What has been demonstrated and what remains openTests on textbook systems such as Lorenz 96 and a reduced-order Kuramoto–Sivashinsky model showed the method closely tracked true outcomes while conventional approaches diverged into impossible values [2][5]. The code is openly available, and a follow-up paper on an energy-constrained operator won the Best Paper Award at the Learning for Dynamics and Control conference [2][5]. However, realistic weather modeling has not yet been tested; the researchers plan to evaluate the method on larger, data-driven atmospheric systems [2][7].## Why this matters for engineering and scienceBounded long-run statistics are essential for climate projections, turbulence modeling in aeronautics, and hydrology [1][6]. By guaranteeing that simulations stay within physically possible bounds, the approach lets practitioners use advanced machine-learning architectures for prediction and control in domains with intrinsic instabilities [1][6]. The open-source release means other groups can integrate the projection layer into their own workflows and test it on the systems they care about most [5][8].
Sources
- New approach enables physically consistent AI models of chaotic systems …
- MIT Framework Keeps AI Models of Chaos Within Physical Bounds
- MIT researchers present method to keep AI models of chaotic systems …
- New Approach Creates Physically Consistent AI Models for Chaos
- MIT researchers develop hard constraints for AI models of chaotic systems
- MIT researchers develop a way to stop runaway predictions in AI physics …
- MIT Launches Physics-Consistent AI for Chaotic Systems
- New approach enables physically consistent AI models of…