DARTS-NetGAB
Research output: · Journal of Mechanical Design
RoleDoctoral researcher and first author
Background
High-fidelity performance models are too slow for immediate design exploration, while standalone surrogate models do not solve the handoff from predicted performance to system geometry. The challenge is to keep simulation, parameter changes, and CAD generation connected.
Method
DARTS-NetGAB integrates ensemble neural-network surrogates, a Panel/Bokeh interface, and ParaturboCAD. A designer can modify system parameters, inspect predicted performance maps, and generate corresponding three-dimensional geometry within one computational workflow.
The project was developed with Joseph Bejjani, Timothy Horvath, Doğukan Üstün, and Jürg Schiffmann at EPFL.
Results
- Reported surrogate errors below 5% for isentropic efficiency and pressure ratio over most of the design domain, increasing near choke conditions.
- Reported performance-map evaluation from approximately one second for a coarse map to 8.5 seconds for a 311,250-point fine map.
- Automated STEP/STL generation through the associated ParaturboCAD workflow.
- Journal publication and conference version documenting the method.
Limitations
Prediction quality depends on the training domain and degrades near difficult operating boundaries. The framework accelerates exploration; it does not replace higher-fidelity analysis or experimental verification for final design decisions.