SC²S Colloquium - Nov 22, 2018

From Sccswiki
Jump to navigation Jump to search
Date: Nov 22, 2018
Room: 00.08.053
Time: 15:00 - 16:00

Benjamin Rüth: Solving the partitioned heat equation using FEniCS and preCICE

This is a preview of his GAMM-CSE 2018 talk

Code reuse is often beneficial from a software engineering perspective. However, reusing code in complex setups that arise in the context of coupled problems is not straightforward. In this talk I want to describe how to solve the partitioned heat equation -- a simple model problem for coupled problems -- using open source software:

Two instances of FEniCS are used to solve the heat equation on originally independent domains. The two domains are non-overlapping and only interacting with each other through a common coupling interface. On the algorithmic level, the two domains are coupled at the interface via Dirichlet-Neumann coupling. The coupling and the communication between the two FEniCS solvers is realized with the open source coupling library preCICE. While FEniCS and preCICE are freely available, the development of glue code (adapter) that allows to couple two instances of FEniCS via preCICE is necessary.

Of course, the aforementioned example is only a proof-of-concept. In the future, we plan to reuse components of this setup to perform more complex simulations -- for example in the area of conjugate heat transfer. Here, the already existing FEniCS solver can be coupled with an appropriate flow solver using preCICE. Additionally, as FEniCS is a general finite element framework, other physical phenomena may be simulated. In this case, parts of the adapter that are specific to the physics have to be modified correspondingly. However, other parts of the adapter that are independent of the physics, such as mesh treatment, can be reused.

Keywords: FEniCS Project, preCICE

KISLAYA RAVI: Neural Network Hyperparameter Optimization using SNOWPAC

This is a Master's thesis introduction advised by Friedrich Menhorn, Hans-Joachim Bungartz and Laura Leal-Taixe

Hyperparameters are crucial factor affecting the accuracy of a neural network. Searching optimal value of hyperparameters without human intervention and guessing has been an important research topic. The aim of this thesis is use SNOWPAC(Stochastic Non-linear Optimizer with Path Augmented Constraints) to automatically search optimal hyperparmeters. SNOWPAC is a stochastic derivative free optimizer which uses surrogate methods to optimize a function. However, some of the hyperparameters of neural network like number of neurons in a layer are an integer. This makes the hyperparameter search a Mixed Integer Non-linear Optimization Problem. Therefore, SNOWPAC is adapted to handle Mixed Integer problems.