feilian-3d

3D U-Net for Urban Wind Speed Prediction Deep learning model for predicting 3D wind speed fields in urban environments from building layout masks. Overview Input: 3D building layout masks (64×512×512 voxels, configurable via env vars) Output: 3D wind speed fields (64×512×512 voxels) Dataset: 539 CFD simulation samples (~40GB) Architecture: 3D U-Net with skip connections Features: Multi-GPU DDP training, Charbonnier loss, physics-informed gradient loss, mixed precision training, OneCycleLR scheduler Model Architecture 3D U-Net: Encoder: 4 downsample levels (MaxPool 2×2×2) Bottleneck: 16× base channels Decoder: 4 upsample levels (TransConv 2×2×2) Skip connections: Concatenate encoder→decoder Parameters: ~45M (base_channels=32) Optimizer: AdamW with OneCycleLR scheduler Warmup: 30% of training (pct_start=0.3) Annealing: Cosine Initial LR: max_lr / 25 Min LR: initial_lr / 10000 Loss: L = Charbonnier + λ·GradientLoss (λ=0.1) Charbonnier loss: More robust to outliers than MSE Gradient loss: Physics-informed loss for spatial smoothness Evaluation Metrics Computed on non-building regions only: MAE: Mean Absolute Error RMSE: Root Mean Squared Error MRE: Mean Relative Error Hit Rate: Fraction within threshold (0.1) NRMSE: Normalized RMSE