Fricke, Clemens David

Wolff, Daniel

Kemmerling, Marco

Elgeti, Stefanie

2023-12-11T13:18:24Z

2023-12-11T13:18:24Z

2023-05-31

<div class="csl-bib-body"> <div class="csl-entry">Fricke, C. D., Wolff, D., Kemmerling, M., &#38; Elgeti, S. (2023, May 31). <i>Enhanced Reinforcement Learning-Based Shape Optimization with Flow Field Information</i> [Conference Presentation]. Math 2 Product (M2P) 2023, Taormina, Italy. http://hdl.handle.net/20.500.12708/190248</div> </div>

http://hdl.handle.net/20.500.12708/190248

Reinforcement Learning (RL) is a subfield of Machine Learning (ML), which contains methods that use trial and error to explore an environment via actions representing the Degrees of Freedom. After every performed action, the environment provides the RL algorithm with information about its state - observation - resulting from the action. During the learning process, additionally, a reward is generated. This reward informs the RL algorithm about the fitness of the applied action for the given environment‘s state. RL is used in many applications, including robotics, control theory, and playing games. Moreover, RL has recently gained traction in the field of design optimization. Viquerat et. al [1] give an overview of multiple areas of fluid mechanics in which RL is currently applied, including the optimization of shapes inside flow fields. Fricke et al. [2] extend the application of RL to the optimization of flow channels in profile extrusion dies, introducing the ReLeSO framework, which enables its users to perform shape optimization via Reinforcement Learning. So far, within shape optimization, the observation returned to the RL algorithm after each action is still limited compared to other applications. Usually, it is restricted to only a few key performance metrics such as flow homogeneity or lift/drag coefficients. In addition, the geometry of the to-be-optimized object can only be included into the learning process via the parameterization used for the optimization. Both limitations can be resolved by providing the full flow field to the RL algorithm as an observation. It is beneficial to not only provide Finite Element solutions but to preprocess the flow field, e.g., by employing a Convolutional Neural Network-based feature extractor. Within our ReLeSO framework, the flow field is converted into an RGB image incorporating up to three flow field variables, which is then provided as an observation to the agent. We apply the proposed method to the second use case presented in [2] and show, that the RL algorithm is capable of learning a suitable strategy given the new observations.

en

Shape Optimization

Reinforcement Learning

Fluid Dynamics

Convolutional Neural Networks

Enhanced Reinforcement Learning-Based Shape Optimization with Flow Field Information

Presentation

Vortrag

RWTH Aachen University, Germany

RWTH Aachen University, Germany

Conference Presentation

C2

C4

X1

Computational Fluid Dynamics

Mathematical and Algorithmic Foundations

Beyond TUW-research foci

30

60

10

E317-01-1 - Forschungsgruppe Numerische Analyse- und Designmethoden

E317-01-2 - Forschungsgruppe Struktur- und Werkstoffsimulation

0000-0002-9495-2285

0000-0003-0141-2050

0000-0002-4474-1666

Math 2 Product (M2P) 2023

RWTH Aachen

RWTH Aachen

thes1136

jara0185

30-05-2023

01-06-2023

On Site

Event for scientific audience

Taormina

IT

Fricke, Clemens David

Maschinenbau

Informatik

Sonstige Technische Wissenschaften

2030

1020

2119

40

30

30

conference paper not in proceedings

Publications

none

en

http://purl.org/coar/resource_type/c_18cp

no Fulltext

E317-01-2 - Forschungsgruppe Struktur- und Werkstoffsimulation

E317-01 - Forschungsbereich Leichtbau

RWTH Aachen University

E317-01 - Forschungsbereich Leichtbau

0000-0002-9495-2285

0000-0003-0141-2050

0000-0002-4474-1666

E317-01 - Forschungsbereich Leichtbau

E317 - Institut für Leichtbau und Struktur-Biomechanik

E317 - Institut für Leichtbau und Struktur-Biomechanik