Session: 37-01 Neural Network based surrogate models and optimization

Paper Number: 101938

101938 - Realtime CFD Based Shape Optimization Using Geometric Deep Learning for Families of Turbomachinery Applications 

Turbomachinery design practices exploit nowadays routinely Computational Fluid Dynamics (CFD) analysis and CFD based shape optimization. In addition to design point conditions, part-load conditions need to be considered when designing a turbomachine. This requires on the one hand CFD analyses with multiple operating points, such as performance curves and maps and on the other hand design space exploration and multi-point optimization studies. The computational cost is further increased if uncertainties are considered within a Robust Design Optimization (RDO) context.

To tackle the computational burden, much effort has been dedicated to accelerating CFD code convergence and to developing efficient optimization strategies including optimization under uncertainty. This has accelerated the turbomachinery design process, increased design productivity, and ultimately saved costs. Despite these efforts, the number of CFD simulations required to conduct these types of analysis remains an important limiting factor in terms of both turnaround time and computational cost. The available time and monetary budgets will likely decide how many operating points are included in the optimization process, how many different designs can be evaluated, how much improvement can be achieved, or how robust the design will be.

Rapid developments in recent years in the field of machine learning and deep learning make it possible to rethink the turbomachinery design process. Realtime predictions of entire performance maps or solution fields for an arbitrary geometry of a given family of applications, such as radial compressors, turbines, or pumps become feasible. Therefore, design productivity can be increased by orders of magnitude.

In this work, a Machine Learning (ML) model based on a geometric convolutional network is applied to an open domain test case, the Rotor 37. The machine learning training data is created based on several different parametric models of the Rotor 37, generating geometrical variations. Then CFD simulations are performed covering several speed-lines within a defined mass flow range. It is demonstrated that the ML model can be trained across geometries originating from different parametric models and predict instantaneously global performance quantities and solution fields with high accuracy.

For such an approach to be applicable with confidence in the turbomachinery design practice, the ML model needs to provide information about its prediction accuracy on a new, unseen geometry to the design engineer. A low prediction accuracy could indicate that the evaluated geometry or operating conditions are outside the predictive range of the ML model. This work shows how a continuous learning approach can be applied to gradually extend this range.

Finally, several optimization studies are performed on top of the trained ML model, which serves as a surrogate of the CFD simulations inside the optimization. In the present work, the key difference to standard surrogate-based design optimization lies in the fact that the trained ML model is independent of the parametric representation and dimensionality of the geometry and can as such be built for a family of applications.

The computational cost associated with these optimization studies is negligible compared to a standard CFD optimization approach. The ML model evaluation for one geometry and operating point is 800 times faster than a CFD simulation, which significantly reduces the turnaround time of an optimization. This allows to include as many operating points as needed in the optimization or to run it several times with different objective and constraint settings, as demonstrated in this work.

It is shown by means of CFD verification of the final optimal geometries that the built ML model is predictive not only for the global quantities typically driving the design optimization, but also for solution fields.

Presenting Author: Alexandre Gouttière Cadence Design Systems Belgiuim

Presenting Author Biography: Alexandre Gouttière holds a master’s degree in mechanical engineering with a specialization in energy fields from the University of Mons in Belgium. He obtained his master’s degree after a 6 month student internship at the TU Delft in the aerospace department. He joined NUMECA in 2019 and work as CFD product engineer for the consulting team which consists in performing simulations, analysis, and optimization response to customer requests in various industrial sectors such as aerospace, aeronautics, turbomachinery, environment, marine or automotive. He joined Cadence with the acquisition of NUMECA in 2021.

Authors:

Alexandre Gouttière Cadence Design Systems Belgiuim
Giacomo Alessi Cadence Design Systems Belgium
Dirk Wunsch Cadence Design Systems Belgium
Luca Zampieri Neural Concept
Charles Hirsch Cadence Design Systems Belgium