Error estimates of physics-informed neural networks for approximating Boltzmann equation

Elie Abdo,Lihui Chai,Ruimeng Hu,Xu Yang

2024-07-11

Abstract:Motivated by the recent successful application of physics-informed neural networks (PINNs) to solve Boltzmann-type equations [S. Jin, Z. Ma, and K. Wu, J. Sci. Comput., 94 (2023), pp. 57], we provide a rigorous error analysis for PINNs in approximating the solution of the Boltzmann equation near a global Maxwellian. The challenge arises from the nonlocal quadratic interaction term defined in the unbounded domain of velocity space. Analyzing this term on an unbounded domain requires the inclusion of a truncation function, which demands delicate analysis techniques. As a generalization of this analysis, we also provide proof of the asymptotic preserving property when using micro-macro decomposition-based neural networks.

Numerical Analysis

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to provide a rigorous error analysis for the performance evaluation of Physics - Informed Neural Networks (PINNs) when approximating the Global Maxwellian near the Boltzmann Equation. Specifically, the paper focuses on the challenges brought by the definition of the non - local quadratic interaction term in the unbounded velocity space domain, which requires the introduction of truncation functions and demands sophisticated analysis techniques. In addition, the paper also provides a proof of the Asymptotic Preserving Property when using neural networks based on micro - macro decomposition.

### Background and Motivation of the Paper

1. **Limitations of Classical Numerical Methods**:
- Classical numerical methods (such as finite difference, finite volume, and finite element methods) have been successful in solving low - dimensional partial differential equations (PDEs), but they have encountered challenges in theoretical analysis and numerical implementation when applied to high - dimensional problems.

2. **Advantages of Deep Learning**:
- Deep neural networks (DNNs) have shown great potential in solving high - dimensional PDEs, especially in terms of robustness and expressive power when dealing with high - dimensional problems.
- As a mesh - free method, DNNs can seamlessly handle complex domains and geometric structures.
- By using the residual error of PDEs as the loss function, DNNs provide a user - friendly implementation method, avoiding the need to construct numerical schemes to approximate derivatives.

### Challenges of the Boltzmann Equation

1. **Curse of Dimensionality**:
- Kinetic Equations describe the evolution of the probability density function of a large number of particles in the phase space, which is usually a six - dimensional problem plus the time dimension.
- When considering uncertainty, the dimension may be even higher.

2. **Non - local Quadratic Interaction Term**:
- The non - local quadratic interaction term in the Boltzmann equation is defined in the unbounded velocity space domain, which requires the introduction of truncation functions and demands sophisticated analysis techniques.

### Main Contributions of the Paper

1. **Rigorous Error Analysis**:
- Provides a rigorous error analysis of PINNs when approximating the Global Maxwellian near the Boltzmann Equation.
- Analyzes the influence of the non - local quadratic interaction term in the unbounded velocity space domain.

2. **Asymptotic Preserving Property**:
- Proves the Asymptotic Preserving Property when using neural networks based on micro - macro decomposition.

### Related Literature

- Recent studies have not only evaluated the computational efficiency and accuracy of PINNs in numerically solving PDEs, but also comprehensively studied their theoretical guarantees.
- Multiple papers have explored in detail the error analysis of PINNs in different problem areas, and these works have provided valuable insights for the current paper.

### Paper Structure

1. **Preliminaries**:
- Provides local estimates of the non - local operators \( K \) and \( \Gamma \).

2. **Error Estimates**:
- Estimates the generalization error and the total error of PINNs when approximating the Boltzmann equation.

3. **Multiscale Boltzmann Equation**:
- Considers the multiscale Boltzmann equation and proves the Asymptotic Preserving Property.

4. **Conclusion**:
- Summarizes the research results and discusses future research directions.

### Main Results

- For any positive time \( T \), assume that \( u \) is a solution of the Boltzmann equation and satisfies certain regularity conditions, then an error estimate can be obtained, indicating that as the loss function converges, the neural network approximate solution will converge to the true solution of the Boltzmann equation.
- Proves that under appropriate conditions, for any small \( \epsilon>0 \), there exists a hyperbolic tangent neural network \( \hat{u} \) such that the generalization error \( E_G[R; \theta]\leq\epsilon \).

### Asymptotic Preserving Property

- Through micro - macro decomposition, an Asymptotic Preserving Neural Network (APNNs) is designed, and in the multiscale Boltzmann equation...