Vortrag "Modelling of microstructures with stochastic geometry and machine learning"

Datum

06.03.2024

Zeit

10:00 - 11:00

Sprecher

Dr. Raimon Tolosano Delgado

Zugehörigkeit

HIF

Sprache

en

Hauptthema

Willkommen

Host

Vinzenz Brendler

Beschreibung

In the geosciences, we often encounter the problem of ensuring representative samples. That is not just about avoiding biases, like in other fields of science, but also about correctly inferring properties on large rock volumes from small samples, often with eye watering contrasts between sample volumes (cm^3-dm^3) and estimation volumes (hm^3-km^3). That is particularly critical for non-additive properties or for properties with complex structures, such as tensor-valued properties, distributional properties, compositional properties or directional properties. Rock microstructure can be measured by a collection of a few such complex structures, including a compositional property (modal mineralogy), a semidefinite matrix-valued property (mineral association) and a few distributional properties (grain size distributions of each relevant minerals): these may include different generations of porosities as “minerals”. Currently, these properties can only be reliably determined by means of scanning electron microscopy based automatic mineralogy systems (SEM-AM) on 2D polished cuts of a few square cm area, in spite of the known stereological degradation that occurs in such 2D measurements with respect to the actual 3D properties. In this paper, a statistical learning methodology that estimates 3D particle microstructures from such images on 2D sections is proposed. The main idea is to make use of a parametric microstructure model, which is then inverted by means of a machine-learning algorithm. In detail, the method works by combining several modelling steps: a fast generator of ore microstructures is implemented using a spectral Gaussian simulation converted into a categorical random field by means of a Markov process. A convolutional neural network (CNN) is then trained to learn the parameters of the microstructure model from 2D sections of 3D realizations. Once the network is trained, it can be used to predict the parameters that characterize the random function, for a given 2D image. In this case, a SEM-AM image is used as input. Thanks to the probabilistic nature of the methods used, the learned parameters allow simulating many different, alternative realizations of 3D volumes compatible with the existing 2D information (not just reconstructing the sample actually measured), hence avoiding at the same time the stereological bias and the volume representativity issues mentioned above. These simulated blocks can then be further used as required, e.g. for simulating the process of crushing of an ore, or a reactive transport of a brine or pollutant flow.

Letztmalig verändert: 06.03.2024, 07:35:55