Hybrid Energy-Based Models for Physical AI: Provably Stable Identification of Port-Hamiltonian Dynamics

A team of researchers has published a significant advance in Physical AI, introducing a new Hybrid Energy-Based Model (EBM) framework designed for the provably stable identification of dynamical systems. The work, led by Simone Betteti and Luca Laurenti, tackles a critical limitation in current AI models for physics and control: the lack of formal stability guarantees. While standard neural networks can be highly expressive, they can also learn unstable behaviors, which is unacceptable for safety-critical applications like robotics, autonomous vehicles, or power grid management. The new hybrid architecture cleverly combines a dynamical visible layer with static hidden layers, expanding the class of models that can be used while ensuring the learned dynamics are inherently stable.

The core innovation is the proof of 'absorbing invariance,' a stability concept that is more flexible than classical global Lyapunov stability but still globally precludes unstable modes. The researchers extended EBM theory to handle non-smooth neural network activations and derived new conditions for stability, exposing a trade-off between expressivity and safety in standard models. Their hybrid framework overcomes this trade-off, and they specifically show these guarantees extend to port-Hamiltonian systems—a fundamental modeling framework in physics and engineering that describes energy flow. Experiments on complex systems like metric-deformed multi-well and ring dynamics validated that the model remains both expressive and safe by design. This work provides a mathematically sound bridge between powerful, data-hungry AI and the rigorous, safety-first world of control theory.