![](https://congressarchive.cimne.com/m2p_2023/assets/a4.png)
Parametric PDEs solvers, uncertainty quantification and parameter estimation
Please login to view abstract download link
This work has been motivated by data assimilation applications in biomedical engineering. In all these applications, a set of partial noisy observations of the system state have to be exploited in order to assess, at best, a set of quantities of interest. We focus here on the applications in which the state evolution can be described by a system of parametric PDEs, and the measurements have to be used in order to estimate the parameters values. Two major difficulties need to be faced. The first one is identifiability: given the observations, can we estimate the parameters? The second one is related to the estimation itself: even when considering linear PDEs, the solution depends potentially in a highly non-linear way on the parameters. When using a variational formulation this often leads to a non-linear non-convex optimisation to be solved, involving a prohibitive computational burden. The goal of the present work is to propose a framework to perform parameter estimation, accounting automatically for identifiability (by using a Bayesian formulation), and for the uncertainties. It can be seen as a possible discretisation of a Bayesian filter. The main idea consists in trying to approximate the system solution as function of space-time and parameters variables. This is done by introducing an efficient low-rank solver. Instead of keeping the discretisation of the parameters domain fixed, we make it evolved based on the system observations. This adaptation task is performed thanks to a Metropolis Hastings algorithm, enabled by a reduced-order interpolator which we can build thanks to the parametric solution. The interplay between the reduced-order interpolator and the parametric solver is crucial to get efficiency: the reduced-order interpolator makes the adaptation fast, getting a more adapted discretisation in parameters contributes to keep the cost of the parametric solver moderate, the parametric solver makes the construction of the ROM interpolator fast. As a result, we can compute, efficiently, mean, standard deviation (other moments or information related quantities) of the posterior distribution of the parameters and the statistics of the system solution. Several test-cases are presented.