Duplicate parameter sets and unexpected LF (VLM) coefficient discrepancy
Hello @yirens ,
I am currently working with the double-delta-aero dataset and noticed that within the 256 training geometries, there are only 240 unique parameter combinations.
For half (8) of these "duplicate" geometries, there are no big issues aside from there being duplicates at all. However, for the remaining half, the surface meshes and high-fidelity (HF) aerodynamic coefficients match perfectly, yet the low-fidelity VLM data inexplicably shows a discrepancy at certain angles of attack (AoA) despite having identical geometric inputs. (In a few rarer cases, the meshes themselves differ slightly while the overall outer mold line remains identical, leading to minor HF discrepancies between 10^−2 and 10^−5 , but the non-negligible LF discrepancy on identical meshes is the primary concern).
Since VLM is typically deterministic for identical geometric inputs, could you provide some insight into why these LF coefficients are splitting? Is this an expected artifact of the compressibility correction, or does it point to a bigger issue with the LF data?
For reference, here are the groups of geomIDs sharing identical parameter combinations that I flagged:
Pairs (2 geomIDs sharing identical parameters):
01100100FQ / 01100100WS
171301-1EW / 171301-1NO
19240101MB / 19240101TK
04160001EV / 04160001TU
141800-3QO / 141800-3ZI
101101-4EP / 101101-4PZ
00240104IW / 00240104TX
04050105CA / 04050105IF
140401-6IU / 140401-6NF
091200-7OI / 091200-7VK
07000107KI / 07000107MF
112301-8AC / 112301-8LP
Triads (3 geomIDs sharing identical parameters):
13010102DH / 13010102LQ / 13010102PI
031301-6JA / 031301-6JK / 031301-6NN
Thank you for your time and for sharing this dataset, and congratulations on graduating.
Hi @k-jani ,
Thank you for the sharp comments. I am aware of certain bugs in the dataset and sorry I was a bit delayed in updating the dataset. I am planning to update the dataset and upload certain scripts for meshing as well.
I think there are three issues you mentioned: 1) duplicated geometry configurations, 2) VLM results discrepancy for identical configuration, and 3) minor discrepancies of the HF CFD meshes between identical geometry. I will go over them one by one.
1) Duplicated geometry configurations
The design of experiment (how geometry configurations are sampled from the design space) is documented in this paper: https://ui.adsabs.harvard.edu/abs/2025arXiv251220941S/abstract
Specifically, Sobol sequence generates nested deterministic space-filling samples that enables the dataset to be used for sensitivity analysis and can be easily expanded. The specific Sobol sequence generator used is SALib. The algorithm is documented in S. Joe and F. Y. Kuo, “Constructing sobol sequences with better two-dimensional projections.” SIAM Journal on Scientific Computing, 30(5):2635-2654, 2008. From the SALib and Joe and Kuo, having duplicated geometries is a known artifact of the Sobol sequence sampling generator, due to a pre-defined base Sobol point list. In practice, skipping the first several base list during the sequence generation can help reduce the duplication. We used a skip value of 16, that leads to 240 unique geometries. Although some geometries are duplicated, it is determined to preserve the original sampling sequence structure to preserve the space filling density, as such property is desired in sensitivity-related applications. However, I do want to thank you again for the detective work and providing this duplication list. I will modify the dataset readme to highlight this feature.
2) VLM results discrepancy for identical configuration
This is a bug. The issue is originated from data curation pipeline. Specifically, when parsing the raw VLM data, and when multi-thread is used, sometimes the physical address conflicts and causing the parser to process the wrong configuration. I am aware of this issue. Thank you so much for providing this bug report and reminding me to update the dataset. I will update the dataset.
3) minor discrepancies of the HF CFD meshes
This is not really a bug nor a desired feature. The meshes used for CFD computation are generated quasi-automatically with mesh scripting. Often, even for identical surface representations, the underlying meshes would not be identical. This cause the minor discrepancies you mentioned. I am planning to upload the meshing configuration script, along with the SUAVE vehicle generation script, so that users can generated their own validation geometries.
Thank you again for the bug report. I will update this dataset. Considering the size of the dataset, it may takes few days to week.
Best,
Yiren