Models. Optuna-tuned MLPs (40 trials each). MF = frozen LF surrogate f
L(x) + HF correction net g([x, f
L]) with residual skip.
Updated 16 Nov 2026 · merged Stage 1 data.
5,639 usable cases from 5,715 delivered; 76 incomplete runs excluded. Counts include held-out tests (HF 212, LF 916).
| HF-test R² (n=212) | CL | CD | CM |
| LF-only | 0.945 | 0.955 | 0.942 |
| HF-only | 0.929 | 0.925 | 0.908 |
| MF | 0.957 | 0.961 | 0.946 |
Surrogate predictions, not CFD. Accuracy drops outside the sampled design space and in regions where CFD often failed (high sweep with high AR).