Dear colleagues,

We are looking for a PhD student for the following project:

For the design of safety critical components and structures, it is vital
to ensure that integrity and reliability are maintained under all
environmental conditions and feasible parameter settings. Physics based
simulation models such as the Finite Element Method provide means to
explore the behavior of a structure for many more scenarios than what
can be measured experimentally. The development of an appropriate model
either on the structural level or for the constitutive behavior of the
materials involved is a complex task. The modelling assumptions can only
be validated based on experimental data, which are either provided by
databases with lab tests under different conditions on the material
scale, or using monitoring techniques for a structural modelling
problem. The challenge is that simulation models usually have multiple
parameters that jointly influence the simulation results , i.e. there is
not a one to one mapping between a single parameter and a single data
set. As a consequence, an inverse problem has to be solved. The goal of
this proposal is to
1. automate the procedure of calibrating model parameters based on
experimental data including objective measures of the model quality
2. identify model discrepancies - which is particularly difficult for
large datasets - by combining using Gaussian process models with physics
based models
3. and suggest model improvements based on a dictionary type of learning
procedure.

All of the procedures are implemented in a Bayesian framework to account
for measurement and model uncertainties. The validation of the developed
methods is initially performed using virtual experiments allowing to
directly prescribe model deficiencies and compare the identified models
to the ground truth. The implementation of the models is done in an open
source framework. As a consequence, the procedures are planned to be
validated in parallel projects on real data sets.

Details:
https://www.bam.de/umantis/EN/952.html

Kind regards,
Philipp Benner