AI Uncovers Patterns in Complex Systems: Implications for Research

Researchers at Duke University have developed a new artificial intelligence framework that reveals simple rules governing complex systems, as detailed in a study published on December 17 in the journal npj Complexity.

This AI framework, inspired by historical dynamicists like Isaac Newton, is capable of analyzing complex, nonlinear systems involving hundreds or thousands of variables, simplifying them into more manageable representations.

Boyuan Chen, director of the General Robotics Lab and Dickinson Family Assistant Professor of Mechanical Engineering and Materials Science at Duke, emphasizes that the ability to distill complex data into simplified rules is crucial for scientific discovery.

The AI combines deep learning with physics constraints to identify essential patterns in time-series data, significantly reducing the dimensionality of models while maintaining accuracy. Testing revealed that the AI consistently identified hidden variables across diverse systems, from pendulum motions to electrical circuits and climate models, creating models that were over ten times smaller than those generated by previous machine-learning techniques.

The results not only enhance predictive capabilities but also improve interpretability, allowing these compact models to connect with established scientific theories. Furthermore, the framework identifies stable states, or attractors, providing insights into system behavior and stability, which is vital for understanding dynamic systems.

Looking ahead, the researchers plan to apply this AI framework to more complex data types and explore its potential to guide experimental design, ultimately contributing to the goal of developing 'machine scientists' that automate scientific discovery.

This research was supported by various programs including the National Science Foundation Graduate Research Fellowship and DARPA initiatives.