Duke University has developed an innovative AI system that can unravel straightforward and comprehensible principles governing intricate systems. This AI delves into the evolution of systems over time, condensing numerous variables into concise equations that effectively represent actual behaviour. The versatility of this method extends to various fields such as physics, engineering, climate science, and biology. Scientists believe it could enhance comprehension of systems where conventional equations are absent or overly intricate.
This novel artificial intelligence framework crafted by Duke University aims to unveil transparent and easy-to-comprehend rules underlying the complexities observed in natural phenomena and modern technologies. Inspired by historical dynamicists who study evolving systems, this system mirrors the approach of Isaac Newton in linking force and motion through equations. It scrutinises data depicting the evolution of complex systems and generates equations that aptly describe their behaviour.
What distinguishes this methodology is its capability to manage complexity beyond human capacity. The AI can simplify nonlinear systems with hundreds or even thousands of interconnected variables into more manageable rules with fewer dimensions.
Boyuan Chen, director of Duke’s General Robotics Lab and the Dickinson Family Assistant Professor of Mechanical Engineering and Materials Science, emphasised the importance of simplifying complex processes for scientific advancement. The new AI framework aligns with mathematician Bernard Koopman’s concept from the 1930s, demonstrating that nonlinear systems can be represented using linear models. The framework leverages deep learning and physics-inspired constraints to streamline complex systems into a smaller set of variables while retaining essential behaviours.
Through analysing time-series data from experiments, this framework identifies significant patterns in how systems evolve. By combining deep learning with physics-based constraints, it distils systems to a reduced set of variables while preserving real-world complexity in a compact model resembling a linear system mathematically.
In testing this approach across various systems like pendulum motion, electrical circuits, climate models, and neural circuits, the AI consistently revealed a small number of hidden variables dictating their behaviour. The resulting models were notably more concise than those generated by previous machine-learning techniques while offering reliable long-term predictions.
Chen highlighted the interpretability alongside accuracy in compact linear models, as it facilitates connecting AI discoveries with established scientific theories developed over centuries. This framework not only predicts but also identifies stable states known as attractors, where systems naturally stabilise over time, aiding in assessing system behaviour.
The researchers emphasise the value of this method when traditional equations are unavailable or too complex to derive. Instead of replacing physics, this approach enhances reasoning using data when physics is obscure or cumbersome to articulate.
Looking ahead, the team plans to explore how this framework can optimise experimental design by selectively collecting data to reveal a system’s structure efficiently. They also aim to apply this method to diverse forms of data like video, audio, and signals from complex biological systems.
This research supports Chen’s General Robotics Lab’s aspiration to develop “machine scientists” for automated scientific discovery. By integrating modern AI with dynamical systems’ mathematical language, this work hints at a future where AI not only recognises patterns but also uncovers fundamental rules shaping both physical and biological realms.