Parameter estimation with model order reduction for elliptic differential equations. Issue 4 (3rd April 2018)

• Record Type: Journal Article
• Authors: Lukassen, Axel Ariaan; Kiehl, Martin
• Is Part Of: Inverse problems in science and engineering. Volume 26:Issue 4(2018)
• Page Start: 479
• Page End: 497
• Publication Date: 2018-04-03
• Subjects: Model order reduction -- parameter estimation -- parametrized partial differential equations -- penalty function -- reduced basis methods -- residual -- thermal block; 65N21; Engineering mathematics -- Periodicals; Inverse problems (Differential equations) -- Periodicals; 620.001515357
• DOI: 10.1080/17415977.2017.1318873
• Languages: English
• ISSNs: 1741-5977

Abstract

Abstract : In this paper a new method for parameter estimation for elliptic partial differential equations is introduced. Parameter estimation includes minimizing an objective function, which is a measure for the difference between the parameter-dependent solution of the differential equation and some given data. We assume, that the given data results in a good approximation of the state of the system.In order to evaluate the objective function the solution of a differential equation has to be computed and hence, a large system of linear equations has to be solved. Minimization methods involve many evaluations of the objective function and therefore, the differential equation has to be solved multiple times. Thus, the computing time for parameter estimation can be large. Model order reduction was developed in order to reduce the computational effort of solving these differential equations multiple times. We use the given approximation of the state of the system as reduced basis and omit computing any snapshots. Therefore, our approach decreases the effort of the offline phase drastically. Furthermore, the dimension of the reduced system is one and thus, is much smaller than the dimension of other approaches. However, the obtained reduced model is a good approximation only close to the given data. Hence, the reduced system can lead to large errors for parameter sets, which correspond to solutions far away from the given approximation of the state of the system. In order to prevent convergence of the parameter estimator to such a local minimizer we penalise the approximation error between the full and the reduced system.