# Mathematical Modelling and Simulation

Working Group 8.41

# Machine learning and data driven modelling

Machine learning (ML) refers to the application of algorithms and statistical models to enable computers to recognize patterns in large amounts of data without having to be explicitly programmed to do so.

In medicine, such approaches are used, for example, to optimize diagnostics, to predict the course of diseases or to automatically evaluate medical images such as X-rays and MRI images.

In the field of numerical flow simulation, ML approaches can be used to replace complex computer simulations with simplified (and therefore computationally less expensive) models. This is particularly relevant for determining the uncertainty of measurement processes and optimizing them since this requires a large number of evaluations of the model

# Machine learning for flow profile prediction

Machine learning methods based on artificial neural networks (ANN) are being applied in an increasing number of applications - also in the field of fluid mechanics. Areas to be mentioned are, for instance, turbulence modelling, interpolation of data from coarser to finer resolution, (active) flow control and the prediction of flows also in their temporal development.
In a TransMeT project in cooperation with working group 3.51 and the company OPTOLUTION, an algorithm based on an ANN is developed that can predict a two-dimensional flow profile of a pipe flow from velocity data along a one-dimensional path. This method will be used, in particular, to reduce the uncertainty of the volumetric flow rate prediction of strongly disturbed velocity profiles - for example due to bends, valves, pipe cross-section changes, etc. - on the basis of path measurements. This method is used in practice for calibration and thus for improving the energy efficiency of drinking water, heating, and cooling networks.

# Deep-Learning based determination of boundary layers in multiphase flows

When gas and liquid flow simultaneously through a pipe, different flow patterns can develop. These patterns mainly differ in the distribution of the fluids in the pipe and the associated formation of structures in the flow. Some of these structures negatively affect the measurement accuracy of flow meters. This influence was investigated as part of the EMPIR project "Multiphase Flow Reference Metrology". This requires detailed characterizations of these flow patterns. For this purpose, a deep-learning model was developed, which extracts the dynamics of the gas-liquid interface in the pipe from high-speed video recordings of the flow, thus enabling classification and characterization. This means that the influence of the flow pattern on the measurement uncertainty can now be further investigated.

# Publications

### 2022

M. Olbrich, L. Riazy, T. Kretz, T. Leonard, D. van Putten, M. Bär, K. Oberleithner;S. Schmelter
International Journal of Multiphase Flow,
2022.

### 2021

T. Mehari;N. Strodthoff
Computers in Biology and Medicine, 141
105114,
2021.
Export as:
BibTeX, XML