Abstract:This paper is about learning the parameter-to-solution map for systems of partial differential equations (PDEs) that depend on a potentially large number of parameters covering all PDE types for which a stable variational formulation (SVF) can be found. A central constituent is the notion of variationally correct residual loss function meaning that its value is always uniformly proportional to the squared solution error in the norm determined by the SVF, hence facilitating rigorous a posteriori accuracy control. It is based on a single variational problem, associated with the family of parameter dependent fiber problems, employing the notion of direct integrals of Hilbert spaces. Since in its original form the loss function is given as a dual test norm of the residual a central objective is to develop equivalent computable expressions. A first critical role is played by hybrid hypothesis classes, whose elements are piecewise polynomial in (low-dimensional) spatio-temporal variables with parameter-dependent coefficients that can be represented, e.g. by neural networks. Second, working with first order SVFs, we distinguish two scenarios: (i) the test space can be chosen as an $L_2$-space (e.g. for elliptic or parabolic problems) so that residuals live in $L_2$ and can be evaluated directly; (ii) when trial and test spaces for the fiber problems (e.g. for transport equations) depend on the parameters, we use ultraweak formulations. In combination with Discontinuous Petrov Galerkin concepts the hybrid format is then instrumental to arrive at variationally correct computable residual loss functions. Our findings are illustrated by numerical experiments representing (i) and (ii), namely elliptic boundary value problems with piecewise constant diffusion coefficients and pure transport equations with parameter dependent convection field.
What problem does this paper attempt to address?
This paper aims to solve the problem of efficient computational exploration of parameter - dependent partial differential equations (PDEs), especially how to construct an approximation method that can efficiently evaluate the parameter - to - solution mapping. Specifically, the paper focuses on the following points:
1. **Problem Background**:
- Parameter - dependent partial differential equation systems are widely present in many applications, and the solutions of these systems usually describe the feasible states of physical systems.
- Researchers often need to select the best parameters to interpret given measurement or observation data. Therefore, it is very important to construct an efficient approximation method to evaluate the mapping from the observation data to the solution manifold.
2. **Research Objectives**:
- The paper proposes a variationally - correct neural residual regression method for learning the parameter - to - solution mapping of parameter - dependent partial differential equation systems.
- The core of this method is to develop a variationally - correct residual loss function, whose value is always in uniform proportion to the square of the solution error, thereby achieving strict posterior accuracy control.
3. **Main Contributions**:
- **Variationally - correct Residual Loss Function**: A variationally - correct residual loss function is proposed, whose value is always in uniform proportion to the square of the solution error, thus ensuring strict control of accuracy.
- **Wide Applicability**: This method is applicable to all PDE types for which a stable variational form (SVF) can be found, including dissipative elliptic or parabolic problems, transport problems, and dispersion models.
- **Computational Feasibility**: By combining the discontinuous Petrov - Galerkin (DPG) method and the mixed hypothesis class, a computable variationally - correct residual loss function is proposed, which solves the optimization problem in high - dimensional parameter spaces.
- **Numerical Experiments**: The effectiveness of the method is verified through numerical experiments, including elliptic boundary value problems and linear transport equations with parameter - dependent convection fields.
4. **Technical Details**:
- **Variationally - correct Loss Function**: The loss function is based on a single variational problem, which is associated with a family of parameter - dependent fiber problems, using the concept of direct integral of Hilbert space.
- **Mixed Hypothesis Class**: A mixed hypothesis class is used, whose elements are piecewise polynomials on low - dimensional space - time variables, with coefficients depending on parameters and can be represented by neural networks.
- **Two Scenarios**:
- **Scenario (S)**: The test space can be chosen as the \( L^2 \) space (e.g., for elliptic or parabolic problems), and in this case, the residual can be directly evaluated in the \( L^2 \) space.
- **Scenario (UW)**: When the trial space and the test space depend on parameters (e.g., for transport equations), the ultra - weak form is used, combined with the DPG method, to achieve a variationally - correct computable residual loss function.
5. **Conclusion**:
- Through the above - mentioned method, the paper provides an efficient and accurate approximation method for parameter - dependent PDE systems, especially suitable for problems in high - dimensional parameter spaces.
In summary, this paper solves the problems of efficient computation and accuracy control in parameter - dependent PDE systems by proposing a variationally - correct neural residual regression method, providing new tools and methods for research in related fields.