Upload folder using huggingface_hub

README.md

@@ -0,0 +1,29 @@
+---
+library_name: pytorch
+tags:
+- tabular
+- pytorch
+- classification
+- aviation
+- wing-design
+datasets:
+- ecopus/transport-wings-500
+license: mit
+---
+
+# Wing Selector MLP
+
+This repository contains a PyTorch MLP that scores aircraft-style wings **within the same airfoil** for a chosen objective:
+- **min_cd** (minimize drag),
+- **max_cl** (maximize lift),
+- **max_ld** (maximize lift-to-drag).
+
+It was trained on the dataset **[ecopus/transport-wings-500](https://huggingface.co/datasets/ecopus/transport-wings-500)**.
+
+## Files
+- `best.pt` – best checkpoint by validation top-1@group  
+- `last.pt` – final checkpoint after training  
+- `config.json` – input dim, #airfoils, feature scaler stats  
+- `feature_names.json` – expected feature order  
+- `airfoil_vocab.json` – airfoil name → id mapping used during training  
+- `inference.py` – minimal loader & scoring helper

airfoil_vocab.json

@@ -0,0 +1,12 @@
+{
+  "atr72sm_fixed": 0,
+  "b707a_fixed": 1,
+  "b707b_fixed": 2,
+  "b707c_fixed": 3,
+  "b707d_fixed": 4,
+  "b707e_fixed": 5,
+  "b737a_fixed": 6,
+  "b737c_fixed": 7,
+  "sc20612_fixed": 8,
+  "sc20714_fixed": 9
+}

best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c518975bc78802090eba638d3a9ffc94b2cba174e08b52cc43de3973f507ea65
+size 86755

config.json

@@ -0,0 +1,58 @@
+{
+  "in_dim": 21,
+  "n_airfoils": 10,
+  "feat_stats": {
+    "means": [
+      2.2974946202834445,
+      0.6872147750606139,
+      0.2578163856764634,
+      0.37480359852313994,
+      2.3189779766400656,
+      9.467880272467932,
+      1.0870784179866313,
+      0.5000662527711445,
+      0.1509065439546441,
+      0.5227734553565582,
+      0.6186037808159988,
+      0.41770886595050494,
+      -1.5537850901646153,
+      1.4478309332036345,
+      -3.993618442217509,
+      1.7111458281675975,
+      0.0070095833313340945,
+      113.27690156300862,
+      6.605367994308471,
+      -1.7461201533675195,
+      1.0
+    ],
+    "stds": [
+      0.4446350828425311,
+      0.13155030403161297,
+      0.07064456938056878,
+      0.07191935098350889,
+      0.632331195206683,
+      2.6987141911651236,
+      0.20895228281937822,
+      0.09680051733887858,
+      0.029613501772416416,
+      0.10120953118257814,
+      0.11874387208280242,
+      0.08241397400630304,
+      0.6453231247887586,
+      0.3816419414156409,
+      1.1807503216075736,
+      0.19466236480543672,
+      0.0007977911059154624,
+      11.519966166557653,
+      0.24553951235854007,
+      1.3978423603895884,
+      1.0
+    ]
+  },
+  "objectives": [
+    "min_cd",
+    "max_cl",
+    "max_ld"
+  ],
+  "framework": "pytorch"
+}

feature_names.json

@@ -0,0 +1,23 @@
+[
+  "span_m",
+  "root_chord_m",
+  "tip_chord_m",
+  "taper",
+  "area_m2",
+  "aspect_ratio",
+  "mac_m",
+  "chord_mean",
+  "chord_std",
+  "chord_mid",
+  "chord_q1",
+  "chord_q3",
+  "twist_mean",
+  "twist_std",
+  "washout_deg",
+  "cl_max",
+  "cd_min",
+  "ld_max",
+  "cla_per_rad",
+  "alpha0l_deg",
+  "has_polar"
+]

inference.py

@@ -0,0 +1,57 @@
+
+import os, json, numpy as np, torch, torch.nn as nn
+
+OBJECTIVES = ["min_cd", "max_cl", "max_ld"]
+
+class MLPSelector(nn.Module):
+    def __init__(self, in_dim:int, n_airfoils:int, obj_dim:int=3, af_embed_dim:int=8, hidden:int=128):
+        super().__init__()
+        self.af_emb = nn.Embedding(n_airfoils, af_embed_dim)
+        self.net = nn.Sequential(
+            nn.Linear(in_dim + obj_dim + af_embed_dim, hidden),
+            nn.ReLU(),
+            nn.Linear(hidden, hidden),
+            nn.ReLU(),
+            nn.Linear(hidden, 1),
+        )
+    def forward(self, x, obj_id, af_id):
+        B = x.size(0)
+        obj_oh = torch.zeros(B, 3, device=x.device)
+        obj_oh[torch.arange(B), obj_id] = 1.0
+        af_e = self.af_emb(af_id)
+        z = torch.cat([x, obj_oh, af_e], dim=1)
+        return self.net(z).squeeze(1)
+
+def load_selector(local_dir=".", device="cpu"):
+    ckpt_path = os.path.join(local_dir, "best.pt")
+    if not os.path.exists(ckpt_path):
+        ckpt_path = os.path.join(local_dir, "last.pt")
+    if not os.path.exists(ckpt_path):
+        raise FileNotFoundError("best.pt/last.pt not found in "+local_dir)
+
+    ckpt = torch.load(ckpt_path, map_location=device)
+    cfg = {
+        "in_dim": int(ckpt["in_dim"]),
+        "n_airfoils": int(ckpt["n_airfoils"]),
+        "feat_stats": {
+            "means": np.array(ckpt["feat_stats"]["means"], dtype=np.float32),
+            "stds":  np.array(ckpt["feat_stats"]["stds"],  dtype=np.float32),
+        }
+    }
+    model = MLPSelector(cfg["in_dim"], cfg["n_airfoils"])
+    model.load_state_dict(ckpt["model"])
+    model.to(device).eval()
+    return model, cfg
+
+def standardize(X_raw: np.ndarray, means: np.ndarray, stds: np.ndarray) -> np.ndarray:
+    X_imp = np.where(np.isfinite(X_raw), X_raw, means)
+    return (X_imp - means) / np.where(stds==0, 1.0, stds)
+
+def score_wings(model, X_std: np.ndarray, airfoil_id: int, objective: str, device="cpu"):
+    obj_id = OBJECTIVES.index(objective)
+    X = torch.tensor(X_std, dtype=torch.float32, device=device)
+    obj_ids = torch.full((X.size(0),), obj_id, dtype=torch.long, device=device)
+    af_ids  = torch.full((X.size(0),), airfoil_id, dtype=torch.long, device=device)
+    with torch.no_grad():
+        probs = torch.sigmoid(model(X, obj_ids, af_ids)).cpu().numpy()
+    return probs  # higher = better

last.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c319f9257b5b4e4b6c2fe803ef769ab5d374c2761a5ff86009e8c24f8c2f52e7
+size 86755

prediction_max_cl_b707b_fixed.json

@@ -0,0 +1,10 @@
+{
+  "objective": "max_cl",
+  "airfoil": "b707b_fixed",
+  "predicted_wing_id": "wing_0373",
+  "pred_prob": 0.9353432059288025,
+  "span_m": 2.4116458892822266,
+  "root_chord_m": 0.911438524723053,
+  "tip_chord_m": 0.4535040259361267,
+  "taper": 0.4975695312023163
+}

prediction_min_cd_b707b_fixed.json

@@ -0,0 +1,10 @@
+{
+  "objective": "min_cd",
+  "airfoil": "b707b_fixed",
+  "predicted_wing_id": "wing_0003",
+  "pred_prob": 0.934077262878418,
+  "span_m": 2.0890963077545166,
+  "root_chord_m": 0.880916953086853,
+  "tip_chord_m": 0.3620271682739258,
+  "taper": 0.41096627712249756
+}