Symplectic Neural Flows for Modeling and Discovery

Hamilton’s equations are fundamental for modeling complex physical systems, where preserving key properties such as energy and momentum is crucial for reliable long-term simulations. Geometric integrators are widely used for this purpose, but neural network-based methods that incorporate these principles remain underexplored. This work introduces SympFlow, a time-dependent symplectic neural network designed using parameterized Hamiltonian flow maps. This design allows for backward error analysis and ensures the preservation of the symplectic structure. SympFlow allows for two key applications: (i) providing a time-continuous symplectic approximation of the exact flow of a Hamiltonian system purely based on the differential equations it satisfies, and (ii) approximating the flow map of an unknown Hamiltonian system relying on trajectory data. We demonstrate the effectiveness of SympFlow on diverse problems, including chaotic and dissipative systems, showing improved energy conservation compared to general-purpose numerical methods and accurate approximations from sparse irregular data. We also provide a thorough theoretical analysis of SympFlow, showing it can approximate the flow of any time-dependent Hamiltonian system, and providing an a-posteriori error estimate in terms of energy conservation.

💡 Research Summary

The paper introduces SympFlow, a time‑dependent symplectic neural network built from parameterized Hamiltonian flow maps. Unlike physics‑informed neural networks (PINNs) or Hamiltonian neural networks (HNNs) that enforce physical laws only through loss terms, SympFlow embeds the symplectic structure directly into its architecture. Each layer corresponds to the exact flow of a Hamiltonian that depends either on position (via a potential V_q(t,q)) or momentum (via V_p(t,p)), mirroring the splitting schemes used for separable Hamiltonians. By composing L such layers, the overall map ψ̄(t,·) remains a Hamiltonian flow; Proposition 1 shows that the composition of two Hamiltonian flows is itself a Hamiltonian flow with an explicit combined Hamiltonian H₃(t,x)=H₂(t,x)+H₁(t,ϕ₂⁻¹(t,x)). This guarantees that ψ̄(0,x)=x and that the Jacobian satisfies the symplectic condition for all t.

Theoretical contributions include Theorem 1, proving that SympFlow is a universal approximator for any time‑dependent Hamiltonian flow, and Theorem 2, which provides an a‑posteriori error bound based on the “shadow Hamiltonian” that can be extracted from a trained model (Equation 7). This enables backward error analysis and quantifies energy drift as |Ĥ(t,ψ̄(t,x₀))−Ĥ(0,x₀)|.

Practically, SympFlow serves two purposes: (i) given a known Hamiltonian system, it yields a continuous, symplectic approximation of the exact flow without resorting to external integrators; (ii) given trajectory data from an unknown system, it learns both the flow map and the underlying Hamiltonian. Experiments cover three benchmark systems: a simple harmonic oscillator, the chaotic Henon‑Heiles system, and a damped harmonic oscillator (a non‑conservative system). For the latter, the authors adopt a phase‑space doubling technique that casts dissipative dynamics into a larger Hamiltonian system, preserving symplecticity while modeling energy loss. Across all tasks, SympFlow outperforms standard Neural ODEs, HNNs, and earlier symplectic networks (e.g., SympNet) in long‑term energy conservation, trajectory fidelity, and data efficiency, especially when training data are sparse or irregularly sampled.

The paper also discusses related work, distinguishing SympFlow from PINNs (which embed constraints in loss functions) and from fixed‑step symplectic neural networks that lack time‑dependence. It highlights that SympFlow’s time‑dependent formulation allows it to handle non‑separable Hamiltonians and to be directly compared with traditional geometric integrators.

In conclusion, SympFlow demonstrates that embedding the symplectic structure at the network level yields models that are both physically faithful and numerically robust. The authors suggest future extensions to high‑dimensional systems, port‑Hamiltonian frameworks, and control‑oriented applications.