Watch a double pendulum swing and you’ll see what looks like randomness made physical. Its metal arms jerk and loop in patterns that seem to mock prediction. Yet every twitch follows rules, ones humans have chased for centuries without quite catching. The math gets ugly fast. Scale up to weather systems or neural firing and even supercomputers choke on the complexity.
A team at Duke University has built an AI that cuts through this mess. Feed it raw measurements from a chaotic system and it spits out compact equations, the kind scientists can actually work with. The framework, published December 17 in npj Complexity, doesn’t just predict what happens next. It reveals why.
The trick lies in finding hidden simplicity. Complex systems often have thousands of interacting variables, but their long-term behavior frequently depends on just a handful. The Duke AI hunts for that small set. In one test, it modeled a nonlinear oscillator using three variables. Previous machine learning methods needed 100 to achieve similar accuracy.
Linear Equations from Nonlinear Chaos
Linear equations hold a special place in science. They’re mathematically clean, they allow long-term predictions, and they connect to centuries of theoretical tools. The problem is that real systems rarely behave linearly. Neural activity, electrical circuits, climate dynamics, they all twist and fold through state space in ways that resist simple description.
The Duke framework uses a physics-informed autoencoder, a type of deep learning architecture that respects the logic of dynamical systems theory. It processes time-series data and searches for coordinates where the system’s evolution becomes linear. The AI also identifies attractors, the stable states where systems eventually settle after disturbance.
“When a linear model is compact, the scientific discovery process can be naturally connected to existing theories and methods that human scientists have developed over millennia. It’s like connecting AI scientists with human scientists,” explains Boyuan Chen, director of Duke’s General Robotics Lab.
The team tested their system on nine different setups. Simple pendulums. Chaotic double pendulums. The Lorenz-96 model, a standard benchmark for weather prediction research. In that last case, involving 40 variables, the AI compressed the essential physics down to 14 dimensions while maintaining accurate forecasts over long time horizons.
What the Neurons Were Hiding
One result stands out. The researchers fed their AI data from the Hodgkin-Huxley model, the Nobel Prize-winning equations that describe how neurons generate electrical signals. The AI found redundancy that wasn’t obvious from the original mathematics. Decades of work by human scientists hadn’t revealed just how much compression was possible.
The framework uses a technique called annealing-based regularization to avoid false modes, patterns that look meaningful but lead nowhere. It starts with simple guesses and refines gradually, distinguishing signal from noise without getting distracted by statistical artifacts.
None of this replaces physics. The researchers are clear on that point. But when equations are unknown, incomplete, or simply too tangled to derive by hand, the AI offers a path forward. It can guide scientists toward the right mathematical structure before theory catches up.
The team plans to extend the approach to video and audio data. The goal isn’t pattern detection alone. It’s uncovering fundamental rules, the kind that let humans reason about systems rather than just simulate them.
npj Complexity: 10.1038/s44260-025-00062-y
ScienceBlog.com has no paywalls, no sponsored content, and no agenda beyond getting the science right. Every story here is written to inform, not to impress an advertiser or push a point of view.
Good science journalism takes time — reading the papers, checking the claims, finding researchers who can put findings in context. We do that work because we think it matters.
If you find this site useful, consider supporting it with a donation. Even a few dollars a month helps keep the coverage independent and free for everyone.