Symplectic Neural Networks are neural network architectures that preserve the symplectic structure of Hamiltonian dynamics — ensuring that the learned dynamics conserve energy and phase-space volume, which is critical for accurate long-term prediction of physical systems.

How Symplectic Networks Work

• Symplectic Structure: Hamiltonian systems preserve the symplectic 2-form $omega = dp wedge dq$.
• Symplectic Integrators: Use integration schemes (leapfrog, Störmer-Verlet) that preserve this structure exactly.
• Network Design: Compose symplectic maps (shear transformations) to build a neural network that is inherently symplectic.
• Separable Hamiltonians: $H(q,p) = T(p) + V(q)$ structure enables efficient symplectic layers.

Why It Matters

• Energy Conservation: Standard neural ODE solvers accumulate energy errors — symplectic networks conserve energy by construction.
• Long-Term Prediction: Symplectic structure ensures bounded errors over long integration times.
• Physics-Informed: Embeds fundamental physics (conservation laws) directly into the architecture.

Symplectic Networks are physics-preserving neural dynamics — architectures that maintain the fundamental conservation laws of Hamiltonian mechanics.