Digital Image Correlation enriched quantification of polymorphic random field uncertainty in finite-element models

The objective of this PhD is to develop an inverse method for the quantification of imprecise random fields, based on full-field strain measurements (Digital Image Correlation (DIC)). A physics-informed deep-learning architecture (e.g., an intelligent combination of a convolutional/recurrent Neural Networks or deep Gaussian process with the governing physical equations) is trained to predict imprecise random-field descriptors (covariance function or power-spectral density), based on a set of displacement and strain fields obtained via DIC, the governing underlying differential equations and prior expert knowledge. To train this physics-informed deep-learning architecture, a predefined set of parametrized random field descriptors is propagated (e.g., by sampling the correlation length in a Gaussian covariance function) to resulting strain fields via a numerical model of the structure and a recently introduced DIC simulator. To limit the training expense of the deep-learning architecture, non-linear manifold learning techniques such as Kernel-PCA, local linear embedding or recently introduced supervised-reduction techniques are used to pre-process the high-dimensional (>100.000 degrees of freedom) strain data and identify a lower-dimensional representation.