Berkeley Fluids Seminar

April 28, 2014

Dr. Alain Demeulenaere (NUMECA-USA, Inc.)

CFD-based design optimisation of turbomachinery blades: Application of genetic algorithms and neural networks

Designing turbomachinery blades is a complex task involving many different objectives and constraints, coming from various disciplines. In order to help the designer in this task, various software codes are now available to define advanced blade geometries (CAD systems), compute flow fields (CFD codes) and the mechanical and thermal stresses inside the blade metal (FEA codes). Although CFD solvers are getting more accurate and faster, they do not provide algorithms to automatically optimize the performance. As a consequence, blade designers often start from an existing geometry and try to improve it based on trial and error procedures. However, the short design time schedules often imposed by the market do not allow the designers to test many modifications and to explore outside of the company's knowledge

Major improvements are expected in terms of reduced design time, reduced engineering time, and also better performance and exploration of new design spaces. This challenge can only be tackled by selecting and further developing efficient design algorithms, integrated into software dedicated to this specific task. Optimization techniques are being utilized more and more used by industry, although limited by timeframes and computer resources.

Several families of optimization strategies exist. The one that will be presented in our talk relies on the interaction between a genetic algorithm, an artificial neural network, a database and user-generated objective functions and constraints. Gradient-based optimization approaches are known to be fast, but may converge toward a local optimum, depending on the selected initial design point. Genetic algorithms offer the advantage of enhancing the probability of reaching the global optimum, but may require thousands of iterations. Their direct coupling with a three-dimensional Navier-Stokes solver cannot be considered under the framework of an industrial design process. The major idea of the methodology employed under FINE^TM/Design3D is that the evaluation of the successive designs is performed using an artificial neural network instead of a flow solver, which permits use of the genetic algorithms in an efficient way. The accuracy of the optimization depends on the knowledge of the neural network, which is fed by design examples stored in a database.Several families of optimization strategies exist. The one that will be presented in our talk relies on the interaction between a genetic algorithm, an artificial neural network, a database and user-generated objective functions and constraints. Gradient-based optimization approaches are known to be fast, but may converge toward a local optimum, depending on the selected initial design point. Genetic algorithms offer the advantage of enhancing the probability of reaching the global optimum, but may require thousands of iterations. Their direct coupling with a three-dimensional Navier-Stokes solver cannot be considered under the framework of an industrial design process. The major idea of the methodology employed under FINE^TM/Design3D is that the evaluation of the successive designs is performed using an artificial neural network instead of a flow solver, which permits use of the genetic algorithms in an efficient way. The accuracy of the optimization depends on the knowledge of the neural network, which is fed by design examples stored in a database.

The talk will demonstrate the application of multipoint and multi-disciplinary optimization strategy to the design of industrial bladings. The major drawback of most published optimization or inverse design techniques is that the optimization is performed at one single operating point, with the danger of observing serious deterioration of the performance at off-design conditions. The generality of the formulation of the FINE^TM/Design3D optimization techniques allows the objective function to be based on the evaluation of the performance at different working conditions. One may either optimize the performance at several given working conditions, or maximize the nominal conditions while imposing that the off-design performance is not deteriorated.

Another very important aspect that is required in order for such techniques to be used industrially relies on the ability to guarantee that the geometries generated will be acceptable mechanically. This can be somehow expected by bringing intuitive constraints on the shapes, but with some level of risk. Full guarantee is possible by bringing a mechanical tool inside the optimization cycle.

Several industrial examples will be shown, illustrating the potential of optimization for improving design performance and widening search ranges.

Back to my webpage

Acknowledgments

Prof. Graham Fleming (Vice Chancellor for Research, UC Berkeley)

Prof. Eliot Quataert on behalf of The Theoretical Astrophysics Center and the Astronomy Department (UC Berkeley)

Prof. Philip S. Marcus on behalf of the Mechanical Engineering Department (UC Berkeley)

Prof. Michael Manga (Earth and Planetary Science, UC Berkeley)

Prof. Evan Variano (Civil and Environmental Engineering, UC Berkeley)