11 ML
Main Components
Download an ONNX wind-prediction model from HuggingFace for the Wind Predictor component. Yel 2.0 is public; Esen 1.0 and Poyraz 1.0 need a HuggingFace token. All are 8-channel Wind Predictor models. (Yel 1.0 is a different architecture — the GAN image model the hosted Wind Predictor (Cloud) component runs via its API — and cannot be loaded here.) Models cache in ~/Eddy3D/Models/ and are reused on subsequent runs (a model already downloaded to the old ~/SUS_LAB/ folder is moved over, not re-fetched). Method: Kastner et al. (2026), SSRN preprint 6401886, doi:10.2139/ssrn.6401886.
Score an ML wind prediction per point: red outside the model's training window, where a feature lies outside the training envelope, or where models disagree by more than 25 %; amber near the window edge, in the envelope's outer 10 %, or where models differ by more than 10 %; green otherwise. The envelope is measured from Dataset Curator CSVs or Verify Prediction logs — none is shipped for the bundled models, because none is published. ML screening prediction, not CFD: use it to compare options, then confirm anything that matters with a Wind Study (planning authorities do not accept surrogate results).
Compare an ML wind prediction with a CFD result at the same points: VDI 3783-9 hit rate (D = 0.25, W = 0.05·U_ref; ≥ 0.66 passes), FAC2, Willmott's index of agreement, R, bias, MAE and RMSE per direction and overall, plus the per-point deviation. Write appends the compared points to a verification log (Dataset Curator columns + U_pred + U_cfd) that is retraining data and a training envelope for Prediction Confidence. ML screening prediction, not CFD: use it to compare options, then confirm anything that matters with a Wind Study (planning authorities do not accept surrogate results).
Run ONNX wind-field prediction end-to-end. Computes SDF, building height, Zrelative, U/Uref, direction features from geometry, assembles the 8-channel input tensor, runs ONNX inference, and outputs predicted wind speeds. Supports legacy 1ch (U), 2ch (U + k) and new 4ch (U + k + Uroof + kroof) models. ML screening prediction, not CFD: use it to compare options, then confirm anything that matters with a Wind Study (planning authorities do not accept surrogate results). Check a result with Prediction Confidence and Verify Prediction. Method: Kastner et al. (2026), SSRN preprint 6401886, doi:10.2139/ssrn.6401886.
Predict a pedestrian wind-speed field from buildings without running CFD, using the hosted Eddy3D model (Yel 1.0, a 512x512 image GAN). Rasterizes the buildings and the analysis plane into the model's input image, sends it to the API, and returns the predicted wind speeds plus a colored result mesh. Runs on Eddy3D's server: needs internet, no GPU and no model download, and the free server may need a minute to wake up. For a local GPU run over arbitrary points and multiple wind directions, use Wind Predictor with a model from ML Model instead. ML screening prediction, not CFD: use it to compare options, then confirm anything that matters with a Wind Study (planning authorities do not accept surrogate results). Method: Kastner & Dogan (2023), Building and Environment 242:110384, doi:10.1016/j.buildenv.2023.110384.
Calculate Pedestrian Wind Comfort using predicted wind fields from the ONNX model. ML screening prediction, not CFD: use it to compare options, then confirm anything that matters with a Wind Study (planning authorities do not accept surrogate results). Method: Kastner et al. (2026), SSRN preprint 6401886, doi:10.2139/ssrn.6401886.
Compute and export wind dataset features from analysis points and building geometry.
Read processed CSV datasets back into Grasshopper. Supports mag_U and all spatial features.
Resample direction-specific wind-magnitude fields onto a new point grid (nearest-neighbour average, with direction-specific rotation). Prepares grids for GAN applications.
Export the solved MRT field as a machine-learning dataset: one row per sensor per hour with spatial features, hourly climate drivers and the MRT/UTCI targets.