Synthetic Data Generation for Deep Learning Models
DS 111: Proceedings of the 32nd Symposium Design for X (DFX2021)
Year: 2021
Editor: Dieter Krause, Kristin Paetzold, Sandro Wartzack
Author: Christoph Petroll, Martin Denk, Jens Holtmannspötter, Kristin Paetzold, Philipp Höfer
Series: DfX
Institution: The Bundeswehr Research Institute for Materials, Fuels and Lubricants (WIWeB)
Section: MBSE
Page(s): 10
DOI number: 10.35199/dfx2021.11
Abstract
The design freedom and functional integration of additive manufacturing is increasingly being implemented in existing products. One of the biggest challenges are competing optimization goals and functions. This leads to multidisciplinary optimization problems which needs to be solved in parallel. To solve this problem, the authors require a synthetic data set to train a deep learning metamodel. The research presented shows how to create a data set with the right quality and quantity. It is discussed what are the requirements for solving an MDO problem with a metamodel taking into account functional and production-specific boundary conditions. A data set of generic designs is then generated and validated. The generation of the generic design proposals is accompanied by a specific product development example of a drone combustion engine.
Keywords: Multidisciplinary Optimization Problem, Synthetic Data, Deep Learning