Learning interpretable and stable dynamical models via mixed-integer Lyapunov-constrained optimization

Researchers Zhe Li and Ilias Mitrai have introduced a novel AI training framework designed to discover stable and interpretable models of dynamical systems directly from data. The core innovation is a mathematical formulation that treats the learning process as a mixed-integer quadratically constrained optimization problem. This approach doesn't just fit a model to the data; it simultaneously searches for the system's governing differential equations and a corresponding Lyapunov function—a mathematical certificate that proves the system's stability. By enforcing Lyapunov stability conditions as hard constraints during training, the method guarantees that the learned model will behave in a predictable, non-chaotic manner around a single equilibrium point, a critical requirement for real-world control applications.

The practical impact is significant for fields like robotics, autonomous systems, and process control. In two case studies, the method successfully discovered the true underlying system model and its stability proof. More importantly, when tested on noisy data—a common real-world challenge—models trained with these Lyapunov constraints achieved measurably higher predictive accuracy compared to standard baseline models that ignored stability. This demonstrates that building physics-informed constraints (like stability) directly into the AI's training objective isn't just about safety; it can also lead to more robust and generalizable models. The work represents a move away from 'black box' AI towards models that are both high-performing and inherently trustworthy by design.