Physics prior residual learning (predict deviation from potential flow Cp)

[tcapelle]

Hypothesis

Predict residual from potential-flow Cp prior to reduce target variance. Lower variance target = easier learning and better generalization.

Implementation

• Compute cp_proxy = -(dsdf_y / dsdf_norm) * is_surface — y-component of unit surface normal (zero on volume nodes)
• Append cp_proxy as input feature at position 25 (after curvature, before Fourier PE) → total input dim 58
• Subtract cp_proxy from normalized pressure target during training
• Add cp_proxy back before denormalization during validation
• fun_dim updated from X_DIM-2+1+32 → X_DIM-2+1+32+1

Results

W&B run: fern/physics-prior-residual (8mxcs36w), best epoch 60, runtime 32 min

Split	mae_surf_p
val_in_dist	18.80
val_ood_cond	14.70
val_tandem_transfer	38.94
val_ood_re	28.60
mean3	24.15

• loss3 = 0.8873
• mean3 = 24.15 vs baseline ~23.88 → REGRESSION (+1.1%)

Analysis

The physics prior does not help on this branch. Key observations:

1. val_tandem_transfer is particularly high (38.94) — for tandem foils, the simple surface-normal prior doesn't capture the hydrodynamic interaction between the two foils. The model must predict a large residual for tandem cases, which likely hurts rather than helps.

2. The dsdf_y signal after normalization may not carry clean directional information since x_mean/x_std normalization changes the scale of both dsdf components differently, distorting the unit normal direction.

3. Single-foil splits improve slightly (in_dist 18.80, ood_cond 14.70) but not enough to compensate for tandem regression.

Recommendation

Close without merging. Potential improvement: compute cp_proxy from the raw (unnormalized) dsdf features, or restrict the prior subtraction to non-tandem samples only.

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

0 out of 2 committers have signed the CLA.
❌ @senpai-advisor
❌ @senpai-fern
senpai-advisor, senpai-fern seem not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}