Master Thesis: Prediction of 3D mass transport properties from 2D data using convolutional neural networks

RISE RESEARCH INSTITUTES OF SWEDEN

Sweden

December 25, 2021

Description

Master Thesis: Prediction of 3D mass transport properties from 2D data using

convolutional neural networks

Background We use both statistical methods and machine learning to understand the relationship between the structure/geometry of porous materials and their mass transport properties, i.e. diffusive transport and fluid flow. In a recent project, we generated a large number of virtual materials structures and computed diffusivity and fluid permeability using lattice Boltzmann methods. The data set consists of 90,000 binary 3D arrays of size 192^3 and the corresponding computed properties, to our knowledge the largest dataset ever of this kind. We used both artificial neural networks (ANNs) and 3D convolutional neural networks (CNNs) to perform nonlinear regression and predict the mass transport properties with high accuracy. However, in many practical cases, only 2D data is available. Therefore, it is of interest to develop methods for prediction of 3D properties using only 2D data, e. g. a single slice of the 3D arrays.

References Prifling et al, 2021, to appear online soon, available on request Röding et al, 2020, https: // www. nature.com/articles/s41598-020-72085-5 (similar earlier study)

Aim The purpose of the proposed MSc project is to develop methods for prediction of 3D mass transport properties from 2D data using 2D CNNs (and/or possibly other machine learning-based regression methods depending on the student's interests).

Project The project consists of implementing and benchmarking appropriate methods for regression/prediction of mass transport properties. The project will give you opportunities to investigate e. g. single-channel and multi-channel 2D CNNs, data augmentation schemes, and hyperparameter optimization. You will benefit from a good knowledge of statistics, spatial statistics, and physics to understand the underlying problem and its relevance. Knowledge of a major deep learning library (preferably TensorFlow/Keras, but it's up to you) is very beneficial. You should be near completion of an MSc program in e. g. applied mathematics/statistics, computer science, or physics. The duration of the project is 20 weeks (30 hp).

Welcome with your application! If this sounds interesting and you would like to know more, please contact Magnus Röding, +46 10 516 66 59. Candidates are encouraged to send in their application as soon as possible but at the latest December 12, 2921. Suitable applicants will be interviewed as applications are received.

Om jobbet Ort

Göteborg

Anställningsform

Visstidsanställning 3-6 månader

Job type

Student - examensarbete/praktik

Kontaktperson

Magnus Röding +46 10 516 66 59

Referensnummer

2021/595

Sista ansökningsdag

2021-12-12

Skicka in din ansökan

Similar Jobs

Inria

France Nov 10, 2021

Add to favorites Read more...

Efficient Neural Rendering: Getting the Best from Points and Volumes

2021-04217 - Efficient Neural Rendering: Getting the Best from Points and About the research centre or Inria department The Inria Sophia Antipolis - Méditerranée center counts 34 research teams as The center's staff (about 500...

ETH Zurich

Switzerland Nov 6, 2021

Add to favorites Read more...

PhD Position – Meshless Simulation of Manufacturing Processes

PhD Position – Meshless Simulation of Manufacturing Processes The Institute of Machine Tools and Manufacturing (IWF) under the Department of Mechanical and Process Engineering (D-MAVT) is an international leading research group on machine tools and...