A Neural Physics Model Now Learns Multiple Stable Resting Points at Once
A new neural network design lets data-driven physical models keep energy structure intact and stay stable around several equilibria at once.
PRO- New neural port-Hamiltonian model preserves stability at multiple equilibria without forcing the Hamiltonian to be convex.
- Architecture multiplies a plain neural net by a smooth step function that vanishes with zero gradient at each equilibrium.
- Prior methods based on input convex neural networks were restricted to a single global equilibrium.
- Theoretical guarantees include local asymptotic stability and an explicit region-of-attraction estimate.
- Beats PH-ICNN baseline on Toda lattice and double pendulum benchmarks in accuracy and correct equilibrium convergence.
- Code available at github.com/EasternBoy/Stability-Preserving-PHS; requires known equilibria up front.
Learning the dynamics of a physical system from data is easy to get wrong in ways that matter. A neural network can fit trajectories beautifully during training and then drift, blow up, or settle at the wrong resting point once you push it outside the sampled region. A new preprint from Binh Nguyen, Nam T. Nguyen, and Truong X. Nghiem tackles that gap for a specific but broadly useful class of systems, proposing a neural port-Hamiltonian model that keeps the physics-based structure and preserves stability at multiple equilibria simultaneously.
Why port-Hamiltonian systems earn their keep
Port-Hamiltonian systems (PHS) provide a framework for modeling, analysis, and control of complex physical systems by emphasizing the flow and storage of energy. In practice, they show up in mechanical systems, electrical circuits, fluid dynamics, robotics, and thermal systems, prized because the formulation bakes in conservation laws and passivity by construction. A port-Hamiltonian neural network (PHNN) tries to inherit those guarantees: by integrating neural networks with port-Hamiltonian structures, PHNNs explicitly encode energy conservation and dissipation principles, enabling robust modeling and prediction of dynamical behaviors and improving stability, interpretability, and generalization.
The catch is that the Hamiltonian, the scalar function that represents the system's stored energy, is usually unknown and has to be learned from state and input traces. Getting it wrong in the wrong place can wreck the long-term behavior of the model.
The convexity trap
Most prior work has enforced stability by parameterizing the learned Hamiltonian with an input convex neural network (ICNN), a design whose output is a convex function of its input. Convexity is convenient because a strictly convex function has exactly one minimum, which conveniently doubles as the system's stable equilibrium. The downside is exactly the same: only one. ICNN-based methods do not apply to many real-world systems whose Hamiltonian is non-convex and possesses multiple isolated stable equilibria.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
