MIT Builds AI That Stays Within Physical Limits For Chaotic Systems

Researchers at MIT have built a new kind of artificial intelligence that can model chaotic systems without drifting into impossible predictions. Chaotic systems include weather patterns and turbulent flows where small changes can grow quickly and make long forecasts unstable. Normal machine learning models often slowly drift away from reality when they run for a long time. The new method adds physics based limits directly into the model so its energy stays bounded. A special projection layer pulls the prediction back when it tries to leave the realistic range. The team tested the idea on classic chaotic examples and kept the rich complex behavior intact. The work was published in the Proceedings of the National Academy of Sciences and shared by MIT. It could make future climate and weather models more trustworthy for long simulations.

image
advertisement
"We will never ask you to login, or register through a third-party advertisement."