Amartya Mukherjee,Melissa M. Stadt,Lena Podina,Mohammad Kohandel,Jun Liu
Abstract:Diffusion models have emerged as a promising class of generative models that
map noisy inputs to realistic images. More recently, they have been employed to
generate solutions to partial differential equations (PDEs). However, they
still struggle with inverse problems in the Laplacian operator, for instance,
the Poisson equation, because the eigenvalues that are large in magnitude
amplify the measurement noise. This paper presents a novel approach for the
inverse and forward solution of PDEs through the use of denoising diffusion
restoration models (DDRM). DDRMs were used in linear inverse problems to
restore original clean signals by exploiting the singular value decomposition
(SVD) of the linear operator. Equivalently, we present an approach to restore
the solution and the parameters in the Poisson equation by exploiting the
eigenvalues and the eigenfunctions of the Laplacian operator. Our results show
that using denoising diffusion restoration significantly improves the
estimation of the solution and parameters. Our research, as a result, pioneers
the integration of diffusion models with the principles of underlying physics
to solve PDEs.
What problem does this paper attempt to address?
The problems that this paper attempts to solve are the forward and inverse problems related to the Laplace Operator, especially the challenges encountered when dealing with the Poisson Equation. Specifically:
1. **Forward problem**: Given a function \( f \in C(\Omega) \), calculate the solution \( u \) that satisfies \( \Delta u = f \) and the boundary condition \( u = 0 \).
2. **Inverse problem**: Given a function \( u \in C^2(\Omega) \) and \( u = 0 \) on the boundary, calculate \( \Delta u \).
### Problem background
The Laplace Operator is a second - order differential operator, defined as \( \Delta=\nabla\cdot\nabla \). It appears in many partial differential equations (PDEs), such as the Poisson Equation, the heat conduction equation and the wave equation. The Laplace Operator is a compact self - adjoint operator, so it has orthonormal eigenfunctions and real eigenvalues.
However, although numerical methods can solve these equations, they will introduce known numerical errors, and for complex PDEs, the computational cost is high. In addition, existing PDE solvers are usually limited to specific types of PDEs.
### Challenges
- **Noise amplification in the forward problem**: In the forward problem, although the Laplace Operator can minimize the measurement noise, in the inverse problem, large eigenvalues will amplify the measurement noise, making parameter estimation difficult.
- **Data scarcity**: The effectiveness and generalization ability of deep learning methods are limited by the scarcity of high - quality training data.
### Solutions
To address these problems, the author proposes a new method - the Denoising Diffusion Restoration Model (DDRM). The specific methods are as follows:
1. **Utilize eigenvalues and eigenfunctions**: By using the eigenvalues and eigenfunctions of the Laplace Operator, project \( u \) and \( f \) onto the shared spectral space, thereby improving the sampling effect.
2. **Improved DDRM algorithm**:
- For the forward problem, sample \( u \) through the conditional distribution \( p(u_{0:T}|f) \).
- For the inverse problem, sample \( f \) through the conditional distribution \( p(f_{0:T}|u) \).
### Experimental results
- **Forward problem**: Using the improved DDRM method, the MAE reached \( 1.175\times 10^{- 6} \), which is significantly better than the result of directly using the DDIM model.
- **Inverse problem**: Using the improved DDRM method, the MAE reached \( 3.215\times 10^{-2} \), which is also significantly better than the result of directly using the DDIM model.
In conclusion, this paper proposes an effective method to solve the forward and inverse problems related to the Laplace Operator by combining the diffusion model and physical principles, which significantly improves the accuracy of parameter estimation.