Leaderboard Submission: CarMRI (PointNet2D + BiLSTM)

model_description.md

@@ -0,0 +1,60 @@
+# CarMRI: Slice-based Surrogate Model for Drag Coefficient Prediction
+
+## Award
+
+**Miltos Petridis Memorial Trophy — Best Student Application Paper** at AI-2025, the 45th SGAI International Conference on Artificial Intelligence, BCS, Cambridge University (December 2025)
+
+## Model Architecture
+
+CarMRI predicts the aerodynamic drag coefficient (Cd) from 3D vehicle point clouds using a sequential, slice-based approach inspired by how MRI/CT scans represent 3D anatomical structures. The architecture (PointNet2D + BiLSTM) has three main components:
+
+1. **PointNet2D (Slice-Level Feature Extractor):** Each 3D point cloud (~100k points) is sliced into S=80 cross-sectional 2D slices along the streamwise (X) axis. Each slice is processed independently by a lightweight 2D adaptation of PointNet using three 1D convolutional layers (channels: 2 -> 32 -> 64 -> 256) with ReLU activations, followed by global max-pooling to produce a 256-dimensional embedding per slice.
+
+2. **Bi-Directional LSTM (Sequence Model):** The sequence of 80 slice embeddings is processed by a 2-layer Bi-LSTM (hidden dimension = 256 per direction) to capture longitudinal geometric dependencies in both flow directions. The final hidden states from both directions are concatenated into a 512-dimensional car-level embedding.
+
+3. **MLP (Regression Head):** A 3-layer fully connected network (512 -> 256 -> 64 -> 1) with ReLU activations and dropout (0.3) regresses the scalar Cd value.
+
+## Implementation Specifics
+
+- **Input Representation:** Each car is represented as a tensor of shape (80, 6500, 2) — 80 slices, each zero-padded to 6500 points with (y, z) coordinates.
+- **Framework:** PyTorch
+- **Total Parameters:** ~2.80 million
+
+## Training Configuration and Hyperparameters
+
+- **Dataset:** DrivAerNet++ (7,713 samples after filtering: 5,398 train / 1,157 validation / 1,158 test)
+- **Data Splits:** Official DrivAerNet++ train/validation/test splits
+- **Loss Function:** Smooth L1 Loss (Huber Loss, beta=1.0)
+- **Optimizer:** Adam (learning rate = 1e-4)
+- **Batch Size:** 4
+- **Epochs:** 100 (best model selected at epoch 68 based on highest validation R²)
+- **Hardware:** Single NVIDIA RTX 4060 GPU (5.9 GB VRAM usage)
+- **Training Time:** ~21 hours
+
+## Results
+
+| Metric | Value |
+|--------|-------|
+| MSE | 6.60 x 10⁻⁵ |
+| MAE | 6.111 x 10⁻³ |
+| Max MAE | 4.50 x 10⁻² |
+| R² | 0.9525 |
+| Inference Latency | 0.025 s/sample |
+| Parameters | 2.80M |
+
+## Link to Paper
+
+- **Springer Nature LNCS:** [https://doi.org/10.1007/978-3-032-11442-6_5](https://doi.org/10.1007/978-3-032-11442-6_5)
+- **arXiv:** [https://arxiv.org/abs/2601.02112](https://arxiv.org/abs/2601.02112)
+
+## Link to Code and Model Weights
+
+- **Repository:** https://github.com/Adarsh-Roy/cd_prediction
+- **Training Code:** https://github.com/Adarsh-Roy/cd_prediction/tree/main/training
+- **Trained Model Weights:** https://github.com/Adarsh-Roy/cd_prediction/tree/main/models
+
+## Authors
+
+- Utkarsh Singh (Delhi Technological University)
+- Adarsh Roy (Indian Institute of Technology, Hauz Khas)
+- Absaar Ali (Delhi Technological University)

test_results.txt

@@ -0,0 +1,5 @@
+Test MSE: 0.0000660
+Test MAE: 0.006111
+Max MAE: 0.0450
+Test R²: 0.9525
+Total inference time: 28.95s for 1158 samples