Topology optimization in turbomachinery is a challenging nonlinear problem with a large number of variables. Compressor efficiency function is dependent on the particular design space and may complicate due to numerical (CFD) solution issue. This makes it difficult to provide high precision, fast convergence, and robustness that is required by a modern, fast paced manufacturing environment. Even the best optimization algorithms exhibit difficulties in addressing these issues. To resolve these difficulties, a new algorithm testing, and training approach was proposed. In the new approach, training function is specially constructed using compressor efficiency function. Efficiency is determined by a combination of analytically derived individual loss functions. To account for some numerical issues associated with automated CFD runs, the obtained efficiency function modified to include solution randomness (noise). After testing most common optimization algorithm, genetic algorithm and pattern search algorithm were selected for training. The training process was automated using nested optimization loops. First, in the inner optimization loop, the best impeller geometry is found. Then in the outer optimization loop, optimum optimization algorithm parameters are found. Comparison of the trained and non-trained algorithm performance demonstrated a positive impact of the new approach. Optimized for compressor efficiency algorithm demonstrated improvement in both convergence precision (over 2% points) and convergence time. Extending the present approach to include other loss mechanisms and CFD generated data is proposed for future work.

This content is only available via PDF.
You do not currently have access to this content.