Abstract:The operator learning has received significant attention in recent years, with the aim of learning a mapping between function spaces. Prior works have proposed deep neural networks (DNNs) for learning such a mapping, enabling the learning of solution operators of partial differential equations (PDEs). However, these works still struggle to learn dynamics that obeys the laws of physics. This paper proposes Energy-consistent Neural Operators (ENOs), a general framework for learning solution operators of PDEs that follows the energy conservation or dissipation law from observed solution trajectories. We introduce a novel penalty function inspired by the energy-based theory of physics for training, in which the energy functional is modeled by another DNN, allowing one to bias the outputs of the DNN-based solution operators to ensure energetic consistency without explicit PDEs. Experiments on multiple physical systems show that ENO outperforms existing DNN models in predicting solutions from data, especially in super-resolution settings.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to learn solution operators that can abide by the physical energy conservation or dissipation laws through data - driven methods in the absence of explicit partial differential equations (PDEs). Specifically, although existing deep neural networks (DNNs) can be used to learn the solution operators of PDEs, they still have difficulties in predicting the dynamics that follow physical laws, especially in cases where the data resolution is low. In addition, some methods require known PDEs for training, which limits their scope of application. To overcome these challenges, the paper proposes a new framework named Energy - consistent Neural Operators (ENOs). The main contributions of ENOs are as follows: 1. **Proposing a learning framework based on energy theory**: This is the first framework to learn the PDEs solution operators of Hamiltonian systems and dissipative systems based on energy theory without explicit PDEs. 2. **Introducing a new penalty function**: This penalty function aims to encourage the time derivatives of DNNs to align with the gradient flow of the energy functional, thereby ensuring that the output of the solution operator complies with the energy conservation or dissipation laws. 3. **Experimental verification**: Experiments on multiple physical systems show that ENOs can better ensure energy consistency while predicting high - resolution solutions, especially when the training data resolution is low. Through these innovations, ENOs can not only predict high - resolution solutions more accurately on low - resolution data, but also ensure that the output of the solution operator complies with physical laws, thereby improving the generalization ability and physical consistency of the model.

Neural Operators Meet Energy-based Theory: Operator Learning for Hamiltonian and Dissipative PDEs

Energy-Dissipative Evolutionary Deep Operator Neural Networks

Disentangled Representation Learning for Parametric Partial Differential Equations

INO: Invariant Neural Operators for Learning Complex Physical Systems with Momentum Conservation

Diffeomorphism Neural Operator for various domains and parameters of partial differential equations

Diffeomorphic Latent Neural Operators for Data-Efficient Learning of Solutions to Partial Differential Equations

Learning Partial Differential Equations with Deep Parallel Neural Operator

Physics-Informed Deep Neural Operator Networks

Deep Energy-Based Modeling of Discrete-Time Physics

Learning in latent spaces improves the predictive accuracy of deep neural operators

Neural Operator: Is data all you need to model the world? An insight into the impact of Physics Informed Machine Learning

Data-Efficient Operator Learning via Unsupervised Pretraining and In-Context Learning

Learning Hamiltonian Neural Koopman Operator and Simultaneously Sustaining and Discovering Conservation Laws

Learning the solution operator of parametric partial differential equations with physics-informed DeepONets

Neural Operator: Learning Maps Between Function Spaces

Pseudo-Differential Neural Operator: Generalized Fourier Neural Operator for Learning Solution Operators of Partial Differential Equations

Operator Learning Enhanced Physics-informed Neural Networks for Solving Partial Differential Equations Characterized by Sharp Solutions

Learning Hamiltonian neural Koopman operator and simultaneously sustaining and discovering conservation law

Nonlinear Operator Learning Using Energy Minimization and MLPs

Physics-informed Discretization-independent Deep Compositional Operator Network

Wavelet Neural Operator for solving parametric partial differential equations in computational mechanics problems