How kth-flowai/fcn-turbulence-predictions-from-wall-quantities works
This repository trains and evaluates fully convolutional networks that reconstruct velocity-fluctuation fields in turbulent channel flow from wall quantities. The main workflow is to read configured TFRecord samples, preprocess inputs and targets, train the FCN, then load a trained model to predict and save test results. WallRecon and OuterRecon are configuration profiles for the prediction problem. The README documents these capabilities; sampled excerpts show the training and evaluation orchestration, but some dataset-processing details are omitted.
Main entry points: train.py Training entrypoint · evaluate.py Evaluation entrypoint · config_sample.py Problem settings · training_utils.py Training utilities
Components 10
Data preparation
- Problem settings
conf/config_sample.py - TFRecord datasets
- Dataset loader
src/tfrecord_utils.py
Reconstruction model
- FCN architecture
models/fcn.py
Training
- Training utilities
src/training_utils.py - Training entrypoint
src/train.py - Saved models
Inference
- Evaluation entrypoint
src/evaluate.py - Prediction results
Other components
- Research user
Connections 17
- Research user → Training entrypoint: starts training
- Research user → Evaluation entrypoint: starts inference
- Problem settings → Training entrypoint: configures
- Problem settings → Evaluation entrypoint: configures
- Problem settings → Dataset loader: sets data options
- Problem settings → FCN architecture: sets model options
- TFRecord datasets → Dataset loader: provides samples
- Training entrypoint → Dataset loader: requests train data
- Evaluation entrypoint → Dataset loader: requests test data
- Training entrypoint → Training utilities: builds model
- Training utilities → FCN architecture: creates FCN
- Training entrypoint → FCN architecture: fits model
- Training entrypoint → Saved models: saves model
- Evaluation entrypoint → Training utilities: loads trained model
- Training utilities → Saved models: reads checkpoint or model
- Evaluation entrypoint → FCN architecture: predicts fields
- Evaluation entrypoint → Prediction results: saves test and predicted fields