Session: 34-05 Physics-based and machine learning models

Paper Number: 151756

Advanced Generative Neural Networks for Predicting Complex 2D Physical Fields With Minimal Data: A VQVAE-Transformer Framework 

This paper introduces a novel generative neural network system for predicting
complex, structured 2D physical fields with a focus on achieving high accuracy,
stability, and computational efficiency. The proposed framework employs a Vec-
tor Quantized Variational Autoencoder (VQVAE) combined with a generative
model that integrates transformer-based structures, inverted ResNet modules,
and mobileSR architectures. The system is specifically designed to operate un-
der constraints typical of complex aerodynamic optimizations, where data avail-
ability is often limited to fewer than 100 samples. It addresses the challenge of
limited data availability by minimizing overfitting and maintaining robustness,
indicating its potential suitability for future use in industrial design and opti-
mization processes.
The capabilities of the proposed system are evaluated through two distinct case
studies. A high-temperature heat pump radial compressor which is the second
stage in a three-stage compression system for super heated steam. The radial
compressor consists of the rotating impeller and the stationary vane geometry.
The second case study is the first stage of a three-stage axial compressor used in
the military sector, where the rotor geometry is varied. These case studies illus-
trate the model’s applicability across different technical domains, highlighting
its potential to serve as a versatile tool in simulation-driven design processes.
We analyze the model’s architecture and provide detailed insights into its capa-
bility to produce physically meaningful predictions, thereby demonstrating its
utility in scenarios that traditionally rely on computationally intensive simula-
tion models.
It is demonstrated that the meta-model is well-suited for performing gradient-
based optimization directly on the surrogate model. This approach allows for
the direct maximization of metrics such as efficiency, but also enables the use
of 2D objective functions, such as minimizing flow separation or targeting spe-
cific local flow phenomena. This type of optimization opens up the possibility of
defining entirely new objective functions that were previously not feasible which
could potentially be integrated into complex design optimization processes in
the future.
In summary, the proposed generative neural network system offers a promising
approach for physical field prediction and optimization, demonstrating strong
potential for adoption in industrial design processes. Its ability to handle low
sample sizes without compromising prediction accuracy makes it a valuable tool
for engineering applications where data scarcity and computational constraints
are prevalent.

Presenting Author: Aryan Karimian German Aerospace Center

Presenting Author Biography: Aryan Karimian obtained his BSc and MSc in Engineering Science at TU Berlin with a focus on fluid mechanics and numerical mathematics. In addition, he acquired experience in the field of turbomachinery, aerodynamics, CFD, and optimization through his studies, student employments as well as his final theses. Currently, he is a research associate at the German Aerospace Center Institute of Propulsion Technology in the field of multidisciplinary compressor optimization and AI methods.

Authors:

Andreas Schmitz German Aerospace Center (DLR)
Robert Schaffrath German Aerospace Center
Christian Voss German Aerospace Center
Aryan Karimian German Aerospace Center
Deeksha Singh German Aerospace Center
Dominik Heinen German Aerospace Center