Search papers, labs, and topics across Lattice.
This paper introduces a reduced-order modeling framework for parameter-dependent PDEs, discretizing the physical domain with finite elements and approximating the parameter dependence separately. They derive error estimates quantifying the interaction between spatial discretization and parameter approximation, achieving algebraic convergence with classical interpolation in low-dimensional parameter spaces. For high-dimensional parameter spaces, they employ extreme learning machine (ELM) surrogates, demonstrating computational savings in quantitative photoacoustic tomography inverse problems while maintaining accuracy.
Extreme Learning Machines offer a computationally efficient alternative to classical interpolation for solving high-dimensional inverse problems governed by PDEs, without sacrificing accuracy.
We develop an interpolation-based reduced-order modeling framework for parameter-dependent partial differential equations arising in control, inverse problems, and uncertainty quantification. The solution is discretized in the physical domain using finite element methods, while the dependence on a finite-dimensional parameter is approximated separately. We establish existence, uniqueness, and regularity of the parametric solution and derive rigorous error estimates that explicitly quantify the interplay between spatial discretization and parameter approximation. In low-dimensional parameter spaces, classical interpolation schemes yield algebraic convergence rates based on Sobolev regularity in the parameter variable. In higher-dimensional parameter spaces, we replace classical interpolation by extreme learning machine (ELM) surrogates and obtain error bounds under explicit approximation and stability assumptions. The proposed framework is applied to inverse problems in quantitative photoacoustic tomography, where we derive potential and parameter reconstruction error estimates and demonstrate substantial computational savings compared to standard approaches, without sacrificing accuracy.
