MIT researchers develop a way to stop runaway predictions in AI physics models
The method guarantees that simulated trajectories stay bounded over arbitrarily long periods. Testing on realistic weather models remains ahead.
Loading page…
The method guarantees that simulated trajectories stay bounded over arbitrarily long periods. Testing on realistic weather models remains ahead.
Listen to this story
The approach gives chaotic-system models a learned energy boundary and a projection layer that forces the modeled energy downward outside that region, providing a theoretical guarantee that trajectories stay bounded over arbitrarily long runs. MIT’s tests found close agreement with physical-system outcomes where conventional methods produced impossible runaway values, but the guarantee is not a promise of long-range forecast accuracy. Weather modeling has not yet been demonstrated; researchers plan to test larger systems, and released code lets others evaluate the method.
Unlike training penalties, the method adjusts predicted system evolution to enforce boundedness on new inputs.
The boundary is learned rather than set as an arbitrary ceiling, leaving the model room to represent complex chaotic behavior.
Tests covered systems from textbook examples to high-dimensional cases; realistic weather modeling remains future work.
A model that looks convincing for a few moments can eventually simulate a world that physics would never allow. MIT researchers have developed a method to prevent one such failure: predictions that grow without limit. Their approach builds an energy-based constraint into AI models of chaotic systems, rather than relying on training alone.
The research, described by MIT News on November 20, appears in the Proceedings of the National Academy of Sciences. Its authors are MIT graduate student Sunbochen Tang and professors Navid Azizan and Themistoklis Sapsis. They address a problem relevant to weather, climate and turbulent flows: preserving a system’s complex behavior without letting the simulation drift into artificial runaway behavior.
Chaotic systems are sensitive to their starting conditions: tiny differences can produce dramatically different outcomes. But that unpredictability does not mean every outcome is physically possible. As Azizan explains, a model can fit a short stretch of data and still fail over longer runs, when small errors accumulate until its simulated values grow without limit.
Previous approaches have often tried to discourage those failures during training. They add penalties or incentives that reward compliance with a desired constraint. Tang says those rules can depend on detailed prior knowledge of the system. More importantly, encouraging a model to obey a rule while it learns does not guarantee that it will obey when given new inputs.
Simply cutting off predictions at a fixed maximum would also stop unlimited growth. But Tang argues that an arbitrary ceiling could prevent a model from learning important, complex features. The team instead lets the model learn an energy-like quantity and a boundary around the region where its chaotic behavior should remain. That boundary comes from learning, rather than a manually chosen cap.
A component called a projection layer then adjusts how the model predicts the system will evolve. Outside the learned boundary, the energy-like quantity must decrease. This draws the simulated system back toward the bounded region. The researchers say enforcing that condition gives them a theoretical guarantee that trajectories remain bounded over arbitrarily long forecast windows.
The design aims to preserve chaos, not smooth it away. Within the learned boundary, the model still has room to represent complex dynamics. Azizan also points to a practical reason for using physics: physical systems often have limited training data, so their governing constraints provide another source of information.
The team tested the method on physical systems ranging from textbook examples to complex systems with many dimensions. MIT reports that its predictions closely matched the true outcomes, while conventional approaches veered into physically impossible predictions. Those results support the method’s promise, but the formal guarantee concerns boundedness—not accurate forecasts arbitrarily far into the future.
Realistic weather modeling remains a planned test, rather than a demonstrated application. The researchers intend to try increasingly large systems. Azizan frames boundedness as one property that can be guaranteed, within a broader effort to make high-stakes AI dependable through more than evidence that it usually works.
The projection layer is modular and, the researchers say, can work with different neural-network architectures. The paper, “Learning dissipative chaotic dynamics with boundedness guarantees,” and open-source code are available online, giving other researchers a way to test the approach on larger problems.
Loading discussion...
Join the conversation
Explain when strong test results should be enough.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.