Abstract:In this paper, we develop an ensemble-based time-stepping algorithm to efficiently find numerical solutions to a group of linear, second-order parabolic partial differential equations (PDEs). Particularly, the PDE models in the group could be subject to different diffusion coefficients, initial conditions, boundary conditions, and body forces. The proposed algorithm leads to a single discrete system for the group with multiple right-hand-side vectors by introducing an ensemble average of the diffusion coefficient functions and using a new semi-implicit time integration method. The system could be solved more efficiently than multiple linear systems with a single right-hand-side vector. We first apply the algorithm to deterministic parabolic PDEs and derive a rigorous error estimate that shows the scheme is first-order accurate in time and is optimally accurate in space. We then extend it to find stochastic solutions of parabolic PDEs with random coefficients and put forth an ensemble-based Monte Carlo method. The effectiveness of the new approach is demonstrated through theoretical analysis. Several numerical experiments are presented to illustrate our theoretical results.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: when numerically simulating partial differential equations (PDEs), how to efficiently find a set of numerical solutions for linear, second - order parabolic partial differential equations. Specifically, these equations may have different diffusion coefficients, initial conditions, boundary conditions and body force terms. The author proposes an ensemble - based time - stepping algorithm. By introducing the ensemble average of the diffusion coefficient function and using a new semi - implicit time integration method, a single discrete system of multiple right - hand - side vectors is combined and solved together, thereby improving computational efficiency.
### Specific problem description
1. **Requirement for multi - parameter simulation**:
In many applications, users need not only a single simulation but also multiple simulations, each with different computational settings, such as different initial conditions, boundary conditions, body force terms and physical parameters. For example, in data assimilation methods in meteorology (such as the ensemble Kalman filter), the numerical weather model is run multiple times by perturbing the initial conditions and uncertain parameters to update the prediction and covariance matrix.
2. **High computational cost**:
Such multi - parameter simulations are very computationally expensive, especially when the forward simulation scale is large. When the traditional implicit time - stepping method is applied to the above - mentioned systems, multiple linear systems need to be solved at each time step, resulting in an increase in storage requirements and computational costs.
3. **Limitations of existing methods**:
Most of the previous methods are only applicable to problems with constant physical parameters, while this paper first proposes an ensemble scheme applicable to non - constant parameter problems and further applies it to PDEs with random coefficients.
### Solution
To solve these problems, the author has developed an ensemble - based time - stepping algorithm, whose main features are as follows:
- **Single discrete system**: By introducing the ensemble average of the diffusion coefficient, all solutions can be obtained at each time step by solving a single linear system with a shared coefficient matrix, thereby reducing storage requirements and computational costs.
- **Wide applicability**: This algorithm is not only applicable to deterministic parabolic PDEs, but also extended to parabolic PDEs with random coefficients.
- **High efficiency**: For large - scale problems, combined with the block Krylov subspace iteration method, the efficiency of multiple numerical simulations can be significantly improved.
### Mathematical expression
Consider the following initial - boundary value problem:
\[
\begin{cases}
u_t-\nabla\cdot [a(x,t)\nabla u]=f(x,t), & \text{in } D\times(0,T), \\
u(x,t)=g(x,t), & \text{on }\partial D\times(0,T), \\
u(x,0)=u_0(x), & \text{in } D,
\end{cases}
\]
where \(D\) is a bounded Lipschitz domain in \(\mathbb{R}^d\), \(d = 1,2,3\), the diffusion coefficient \(a(x,t)\in L^2(W^{1,\infty}(D);0,T)\), the body force term \(f(x,t)\in L^2(H^{-1}(D);0,T)\), the initial condition \(u_0(x)\in H^1_0(D)\cap H^{l + 1}(D)\) and \(l\geq1\).
For an ensemble simulation containing \(J\) independent members, the \(j\)-th member satisfies:
\[
\begin{cases}
u_{j,t}-\nabla\cdot [a_j(x,t)\nabla u_j]=f_j(x,t), & \text{in } D\times(0,T), \\
u_j(x,t)=g_j(x,t), & \text{on }\partial D\times(0,T), \\
u_j(x,0)=u_{0,j}(x), & \text{in } D,
\end{cases}
\]
To improve efficiency, the author proposes a new ensemble - based time - stepping scheme. Define the ensemble average of the diffusion coefficient function at time \(t_n\) as:
\[
a^