The task
A turbo compressor wheel is defined by dozens of geometric parameters. Every design change went from CAD to CFD simulation or a physical prototype, often dozens of times. A CFD run takes days to prepare and days of expensive computation. A prototype takes several weeks to manufacture and test. Between iterations, engineers had no feedback except their own experience.
What we built
Explaining. Statistical analysis (ANOVA, CHAID, Pearson and Spearman correlation) and explanations from trained models show which parameters drive performance, and how strongly.
Predicting. Deep neural networks predict efficiency, air flow and speed directly from the CAD geometry. One prediction covers the whole operating range, so engineers get the full compressor map for a new design at once.
Uncertainty and stress. In the next phase, the models were extended to compressor stress and to the uncertainty of each prediction, with automated model tuning.
The result
- Compressor efficiency predicted with precision comparable to CFD, in seconds instead of days to weeks
- Many design experiments in one day instead of one every few days
- Several dependencies that engineers did not know about were discovered
- The approach carries over to other areas of product design