Abstract:Partial differential equations (PDEs) are commonly derived based on empirical observations. However, recent advances of technology enable us to collect and store massive amount of data, which offers new opportunities for data-driven discovery of PDEs. In this paper, we propose a new deep neural network, called PDE-Net 2.0, to discover (time-dependent) PDEs from observed dynamic data with minor prior knowledge on the underlying mechanism that drives the dynamics. The design of PDE-Net 2.0 is based on our earlier work \cite{Long2018PDE} where the original version of PDE-Net was proposed. PDE-Net 2.0 is a combination of numerical approximation of differential operators by convolutions and a symbolic multi-layer neural network for model recovery. Comparing with existing approaches, PDE-Net 2.0 has the most flexibility and expressive power by learning both differential operators and the nonlinear response function of the underlying PDE model. Numerical experiments show that the PDE-Net 2.0 has the potential to uncover the hidden PDE of the observed dynamics, and predict the dynamical behavior for a relatively long time, even in a noisy environment.

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is: **Automatically discover partial differential equation (PDE) models from observed dynamic data, and be able to make accurate long - term predictions with only a small amount of prior knowledge of the driving dynamic mechanisms**. Specifically, the author proposes a new deep neural network named PDE - Net 2.0, which aims to reveal the PDE models hidden behind complex dynamic data by combining numerical approximation and symbolic multi - layer neural networks, and can predict future behaviors for a long time. ### Description of the Main Problems in the Paper 1. **Limitations of Traditional Methods**: - Traditional PDEs are usually derived mathematically or physically based on some basic principles. However, for complex systems in many modern applications (such as multiphase flow, neuroscience, finance, bioscience, etc.), the mechanisms behind them are still unclear. Therefore, the governing equations of these systems are usually obtained through empirical formulas. - With the rapid development of sensor technology, computing power, and data storage capacity, we can now easily collect, store, and process large amounts of data. This provides new opportunities for discovering new physical laws from data. 2. **Deficiencies of Existing Work**: - Existing methods such as SINDy need to construct a large enough dictionary, which may lead to high memory load and computational cost, especially when the number of model variables is large. - Most of the existing methods process the spatial and temporal information of data separately and fail to fully utilize the time - dependence of PDE models. - Some methods assume that the form of PDE is known, only need to learn a few scalar parameters, or are specifically targeted at specific types of equations (such as convection - diffusion equations), and are difficult to extend to other types of equations. 3. **Objectives of This Paper**: - Design a transparent deep neural network that can reveal the hidden PDE models from the observed complex dynamic data with only a small amount of prior knowledge of the dynamic mechanisms, and at the same time make accurate predictions. - Emphasize the importance of model recovery and prediction, because accurate long - term prediction is an important indicator for verifying the accuracy of the learned PDE model. The trained neural network can be directly applied to predictions under different initial conditions without retraining. ### Innovations of PDE - Net 2.0 - **Flexibility and Expressiveness**: PDE - Net 2.0 not only learns differential operators but also learns nonlinear response functions, having the greatest flexibility and expressiveness. - **Symbolic Neural Network**: Introduce a symbolic neural network (SymNet) to approximate nonlinear response functions, making it possible to reveal the analytical form of PDEs with less prior knowledge. - **Constraints**: By imposing appropriate constraints on convolutional filters, make PDE - Net 2.0 more flexible and be able to better approximate unknown dynamics and make longer - term predictions. In summary, this paper aims to solve the problem of automatically discovering PDE models from data and making long - term predictions through PDE - Net 2.0, especially in the case of limited understanding of the system mechanisms.

PDE-Net 2.0: Learning PDEs from Data with A Numeric-Symbolic Hybrid Deep Network

PDE-Net: Learning PDEs from Data

Neural-PDE: a RNN Based Neural Network for Solving Time Dependent PDEs

An Efficient Method for Solving Two-Dimensional Partial Differential Equations with the Deep Operator Network

DL-PDE: Deep-learning based data-driven discovery of partial differential equations from discrete and noisy data

P^2C^2Net: PDE-Preserved Coarse Correction Network for Efficient Prediction of Spatiotemporal Dynamics

P$^2$C$^2$Net: PDE-Preserved Coarse Correction Network for efficient prediction of spatiotemporal dynamics

SineNet: Learning Temporal Dynamics in Time-Dependent Partial Differential Equations

Deep Latent Regularity Network for Modeling Stochastic Partial Differential Equations

PDE-READ: Human-readable partial differential equation discovery using deep learning

A numerical method to solve PDE through PINN based on ODENet

Trans-Net: A transferable pretrained neural networks based on temporal domain decomposition for solving partial differential equations

NeuralPDE: Modelling Dynamical Systems from Data

Deep-learning based discovery of partial differential equations in integral form from sparse and noisy data

Physical Asymptotic-Solution Nets: Physics-driven Neural Networks Solve Seepage Equations As Traditional Numerical Solution Behaves

Finite Difference Nets: A Deep Recurrent Framework for Solving Evolution PDEs

DISCOVER: Deep identification of symbolically concise open-form PDEs via enhanced reinforcement-learning

Learning to simulate partially known spatio-temporal dynamics with trainable difference operators

An improved data-free surrogate model for solving partial differential equations using deep neural networks

SelectNet: Self-paced learning for high-dimensional partial differential equations