NONSEPARABLE SYMPLECTIC NEURAL NETWORKS

Abstract

Predicting the behaviors of Hamiltonian systems has been drawing increasing attention in scientific machine learning. However, the vast majority of the literature was focused on predicting separable Hamiltonian systems with their kinematic and potential energy terms being explicitly decoupled while building data-driven paradigms to predict nonseparable Hamiltonian systems that are ubiquitous in fluid dynamics and quantum mechanics were rarely explored. The main computational challenge lies in the effective embedding of symplectic priors to describe the inherently coupled evolution of position and momentum, which typically exhibits intricate dynamics. To solve the problem, we propose a novel neural network architecture, Nonseparable Symplectic Neural Networks (NSSNNs), to uncover and embed the symplectic structure of a nonseparable Hamiltonian system from limited observation data. The enabling mechanics of our approach is an augmented symplectic time integrator to decouple the position and momentum energy terms and facilitate their evolution. We demonstrated the efficacy and versatility of our method by predicting a wide range of Hamiltonian systems, both separable and nonseparable, including chaotic vortical flows. We showed the unique computational merits of our approach to yield long-term, accurate, and robust predictions for large-scale Hamiltonian systems by rigorously enforcing symplectomorphism.

1. INTRODUCTION

A Hamiltonian dynamic system refers to a formalism for modeling a physical system exhibiting some specific form of energy conservation during its temporal evolution. A typical example is a pendulum whose total energy (referred to as the system's Hamiltonian) is conserved as a temporally invariant sum of its kinematic energy and potential energy. Mathematically, such energy conservation indicates a specific geometric structure underpinning its time integration, named as a symplectic structure, which further spawns a wide range of numerical time integrators to model Hamiltonian systems. These symplectic time integrators have proven their effectiveness in simulating a variety of energy-conserving dynamics when Hamiltonian expressions are known as a prior. Examples encompass applications in plasma physics (Morrison, 2005) , electromagnetics (Li et al., 2019) , fluid mechanics (Salmon, 1988) , and celestial mechanics (Saari & Xia, 1996) , to name a few. On another front, the emergence of the various machine learning paradigms with their particular focus on uncovering the hidden invariant quantities and their evolutionary structures enable a faithful prediction of Hamiltonian dynamics without knowing its analytical energy expression beforehand. The key mechanics underpinning these learning models lie in a proper embedding of the strong mathematical inductive priors to ensure Hamiltonian conservation in a neural network data flow. Typically, such priors are realized in a variational way or a structured way. For example, in Greydanus et al. (2019) , the Hamiltonian conservation is encoded in the loss function. This category of methods does not assume any combinatorial pattern of the energy term and therefore relies on the inherent expressiveness of neural networks to distill the Hamiltonian structure from abundant training datasets (Choudhary et al., 2019) . Another category of Hamiltonian networks, which we refer to as structured approaches, implements the conservation law indirectly by embedding a symplectic time integrator (DiPietro et al., 2020; Tong et al., 2020; Chen et al., 2020) or composition of linear, activation, and gradient modules (Jin et al., 2020) into the network architecture.

