Designing Poisson Integrators Through Machine Learning

Introduction

The paper discusses the importance of Hamiltonian systems in science and engineering, and the role of geometry in understanding their dynamical behavior. While symplectic geometry has been widely used for developing accurate numerical simulations (symplectic integrators) of Hamiltonian systems, the more general Poisson geometry poses challenges due to its irregular nature. The authors note a lack of general methods for obtaining Poisson integrators, although some attempts exist for linear and integrable Poisson structures.

The paper aims to combine differential geometry with machine learning techniques to develop Poisson integrators. It contributes a geometric setting for describing Poisson geometric integrators on integrable Poisson manifolds, a method for approximating the Hamilton-Jacobi equation using machine learning techniques, and an illustration using the rigid body example.

The authors highlight recent efforts in using symplectic groupoids and symplectic realizations to generate Poisson integrators for linear and integrable cases. Additionally, they mention the use of Poisson structure in learning methods and the design of Poisson neural networks for constant rank Poisson Hamiltonian systems based on the Darboux-Lie theorem.

A Quick Review of Geometry

The provided section introduces some basic concepts in symplectic and Poisson geometry that will be used throughout the paper. It defines symplectic manifolds, which consist of a manifold and a non-degenerate closed two-form on it. It gives the example of the cotangent bundle of a manifold equipped with the canonical symplectic form. Lagrangian submanifolds, which are submanifolds on which the symplectic form vanishes, are also defined.

The section then moves on to Poisson manifolds, which are manifolds endowed with a specific type of bivector field called the Poisson bivector. Casimir functions, which are functions that Poisson commute with all other functions, are introduced. Poisson mappings between Poisson manifolds, which preserve or reverse the Poisson structures, are defined.

Hamiltonian vector fields associated to functions (Hamiltonians) on Poisson manifolds are described. Their flows are Poisson diffeomorphisms.

The motivation for symplectic realizations of Poisson manifolds is given - they allow techniques from symplectic geometry to be used to study the singular Poisson structures. Symplectic groupoids, which are groupoids equipped with symplectic structures satisfying certain conditions, are defined as a way to obtain such symplectic realizations locally.

Finally, an example of a symplectic groupoid integrating a Poisson Lie algebra is provided.

Lagrangian Bisections and Poisson Diffeomorphisms: The Hamilton-Jacobi Equation

The paper discusses methods for approximating the Hamiltonian flow, which describes the time evolution of a dynamical system. The key ideas are:

The authors consider a Poisson manifold and assume there exists a symplectic groupoid associated with it. They observe that Poisson diffeomorphisms on the Poisson manifold can be described through Lagrangian submanifolds in the symplectic groupoid.

The paper cites a previous result showing that if a Lagrangian submanifold satisfies certain conditions, it induces a Poisson diffeomorphism on the original Poisson manifold. This motivates characterizing the Lagrangian submanifold corresponding to the Hamiltonian flow of interest.

The characterization leads to a geometric version of the Hamilton-Jacobi equation, a partial differential equation (PDE) whose solution describes the Hamiltonian flow.

Since solving the PDE analytically is challenging, the paper proposes an approximation approach using machine learning techniques. It involves parametrizing candidate solutions with a neural network, evaluating the PDE residual on a set of points, and minimizing the mean squared residual to find the neural network weights that best approximate the PDE solution.

The optimization leverages backpropagation to compute gradients and modern algorithms like stochastic gradient descent for efficiency. This machine learning-based strategy provides an alternative to classical approximation methods like Taylor expansions.

Numerical Simulations

The paper applies the described methodology to the rigid body system, a benchmark for studying Poisson dynamics. The system is defined on the Lie-Poisson structure of the special orthogonal group SO(3), with a specific Hamiltonian involving the squared coordinates.

To illustrate the methodology, the authors uniformly sampled 80,000 points around the origin, with coordinates ranging from -3 to 3. They used the ADAM optimization algorithm with a neural network architecture to approximately solve an equation related to the system.

The neural network had four layers with 500, 250, 250, and 250 neurons, respectively, and used the hyperbolic tangent activation function. The weights were initialized using Xavier initialization.

Two different initial conditions were chosen to showcase the outcomes of the obtained integrators. The simulated trajectories exhibited qualitative behavior similar to the original system, with good conservation of the Hamiltonian and Casimir functions, even for long simulations.

The evolution of the Hamiltonian, Casimir, and the difference from the original dynamics are plotted for a long trajectory spanning 10,000 iterations with a step size of 0.1.

Conclusions and Future Work

The paper discusses two potential research directions for the geometric framework introduced in a previous work (Vaquero et al., 2023).

Firstly, the authors aim to extend the framework to any Poisson manifold by approximating the symplectic groupoid when it is not readily available. Two main constructions proposed by Cabrera (2022) will be explored to produce local symplectic groupoids that can integrate any Poisson manifold locally.

Secondly, the authors propose a blended approach that combines the Hamilton-Jacobi equation with available trajectory data. This approach is inspired by Physics-Informed Neural Networks (PINNs) described by Raissi et al. (2019). The objective is to find the Lagrangian submanifold that minimizes an objective function composed of two terms: one term ensures the Hamilton-Jacobi equation is satisfied, and the other term ensures the Lagrangian submanifold induces a Poisson transformation that matches the given data.