Project Details
Projekt
Shape optimizations using AI agents to accelerate numerical flow field simulation and to control the island model for massive parallelization
Applicant
Professor Dr.-Ing. Stefan Riedelbauch
Subject Area
Hydraulic and Turbo Engines and Piston Engines
Fluid Mechanics
Fluid Mechanics
Term
since 2022
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 501932169
The overall goal of the project is to extend the automated design process for complex fluid mechanical components. To accelerate the design process with the general design system developed at the applying Institute, it will be coupled with the island model to parallelize the optimization. The island model subdivides the process into serial optimization runs on several interacting islands. In contrast to the current state of the art where artificial intelligence is only used for surrogate models to reduce computational effort, this project aims to extend the entire automated process by using artificial intelligence agents. A data-driven trained agent adds value, for example, in initializing the optimization process, evaluating the fitness function of a candidate solution, or deciding which computational grid is most appropriate for a specific geometry. For example, the creation of a computational grid can be viewed as a mostly manual analysis of the geometry to obtain a well-suited grid that is transferred to a flow solver. Another example is the creation of the parameterization of a rather complex turbomachine, typically processed manually. Both examples are based solely on empirical data. The rule-based decisions are now replaced or supported by data-driven agents. In addition, the agents control the evaluations of the fitness functions and also the behavior in multi objective optimizations. In this sense, the agents extend the flexibility in combination with the island model. Existing data from successful runs are available to evaluate the developed design agents and their performance in the optimization process. The overall work program focuses on two main aspects. At first, the combination of artificial intelligence with the control of the optimization is addressed. The idea is to replace formalized and rule-based aims with data-driven aims. On the other hand, in an abstract sense, the project addresses flexible coupling between programs by transferring knowledge from simple test cases to complex engineering problems using artificial intelligence-based agents. Also, the simulation setup, mesh generation, and evaluation of individuals is accelerated by agents. In particular, agents improve the coupling between the flow solver and the mesher by searching for the best mesh and thus the optimum. By using an unusual kinetic turbine for benchmarking, a strong connection to an engineering problem is established. It also serves as a validation case to evaluate the performance of optimization with all developed agents compared to optimization without them.
DFG Programme
Priority Programmes
Co-Investigator
Dr.-Ing. Alexander Tismer
