SC²S Colloquium - Sep 19, 2018
|Date:||Sep 19, 2018|
|Time:||15:00 - 16:00|
Yehor Yudin: Artificial Neural Network Approach to Design Optimization
This is an external Master's Thesis submission talk advised by Stefan Gavranovic and Sergey Zakharov (Siemens CT)
Design Optimization is a perspective engineering approach which can find solutions for many engineering problems under multiple constraints in a non-analytic manner. The existing algorithms are computationally expensive and could be accelerated by finding approximate results using data-driven methods. This work proposes an approach to predict the material layout of elements with minimal compliance using a convolutional neural network model. The work proposes the corresponding CNN architecture and analyzes the ability of the model to interpolate various parameters and its capacity to be trained on small datasets. Both 2D and 3D cases are considered and an approach to use the model-produced results as an educated guess for further TO performance improvement is proposed and examined.
Keywords: Convolutional Neural Network, Design Optimization
Thomas Bellebaum: Evaluation of different time-synchronization methods for the combination technique
This is an Bachelor's Thesis submission talk advised by Michael Obersteiner. (will be held in German )
Solving partial differential equations (PDEs) numerically via discretization on regular grids in high dimensions is computationally demanding. The sparse grid combination technique is known to reduce the needed work significantly by combining several smaller grids with different mesh widths. Due to the CFL-restriction, to allow for conver- gence using an explicit solving method, timesteps may only be chosen based on the spacial mesh widths of the grids. This thesis aims to exploit the CFL-condition to maximize the timesteps taken and achieve a minimal asymtotic computational complexity for the integration.
Keywords: Sparse grids, time stepping