MIT researchers publish a way to keep chaotic models bounded

The method adds a physics-based constraint to neural networks, addressing simulations that can otherwise grow without limit. The team plans to test it on realistic weather models.
In this article
Researchers at the Massachusetts Institute of Technology have published a method that places a mathematical limit on how models of chaotic systems evolve. Their study in the Proceedings of the National Academy of Sciences describes a component called a dissipative projection layer, which they say guarantees that the learned model’s predictions remain bounded.
In chaotic systems, small differences in starting conditions can lead to very different outcomes. Small model errors can also accumulate during a long simulation, producing values that grow without limit even when the real system stays within bounds, MIT associate professor Navid Azizan said. The study first appeared as an arXiv preprint on 1 October 2024.
A constraint learned from data
The model learns an energy-like function and a boundary alongside its representation of how a system changes. The projection layer adjusts that predicted change so the energy-like quantity must decrease outside the boundary, drawing the simulation toward a bounded region.
That rule is a hard constraint rather than a penalty applied during training. Sunbochen Tang, an MIT doctoral student in aeronautics and astronautics and its Laboratory for Information and Decision Systems, said a training penalty cannot guarantee that a model will obey the intended limits on new inputs. An arbitrary cap can prevent runaway values, he said, but can also restrict complex behavior the model should learn from data.
The researchers say their proof guarantees that every trajectory of the learned system stays bounded and converges to a learned region. Boundedness is necessary for reliable long-run statistics, but it does not guarantee an accurate individual forecast or correct statistics.
Tests and the next step
The team tested the method on the Lorenz 63 and Lorenz 96 systems and a simplified Kuramoto–Sivashinsky model. The researchers said their constrained model stayed bounded and preserved the systems’ long-run statistical patterns, while a neural network without the constraint produced trajectories that grew without limit in finite time.
Tang and Azizan wrote the paper with Themistoklis Sapsis, an MIT mechanical engineering professor and laboratory affiliate. Azizan, a principal investigator at the laboratory, said physics-based constraints can provide another source of information when data on a physical system is limited.
A separate follow-up paper, ECO, by Andrea Goertzen, Tang and Azizan won the Best Paper Award at the 2026 Learning for Dynamics and Control conference. The award was presented on 19 June.
The paper and open-source code are available online. Azizan said the projection layer can work with different neural-network designs; Tang said the team plans to test the approach on data-driven models of realistic weather systems.
Follow AveNews
Earlier in Technology
SuperintelligenceDragos finds evidence of breaches at three superintelligence hosts
Researcher HD Moore says attackers used some credentials planted in an experiment involving nearly two million prompts, while some requests marked for privacy were also exposed. The findings point to risks as prompts pass through multiple providers.
4 min
Chips & ComputingMexican mobile users report lost signal and failed calls
Telcel says service was intermittent in some localities and users were reconnecting, while AT&T Mexico says its network was operating normally.
2 min
SuperintelligenceUS House committees urge limits on Chinese superintelligence models
Their investigation identifies uses at Airbnb, DoorDash and Cursor maker Anysphere, and recommends restrictions for sensitive systems; the proposals are not adopted rules.
4 min
Sources
Show all 4 sourcesShow fewer
MIT News AINew approach enables physically consistent AI models of chaotic systems (opens in a new tab)
- tangsun.github.ioLearning Dissipative Chaotic Dynamics with Stability Guarantees — Sunbochen Tang, Themistoklis Sapsis, Navid Azizan | Sunbochen Tang (opens in a new tab)
- arxiv.org[2410.00976] Learning Dissipative Chaotic Dynamics with Boundedness Guarantees (opens in a new tab)
- lids-2024-development.webflow.ioLIDS researchers win 2026 L4DC Best Paper Award (opens in a new tab)
