WavyChannel-RF-ThermoHydraulics

Model Introduction

This package reproduces the fixed 80/20 Random Forest experiment for a partially porous wavy-channel CFD surrogate. It predicts mean Nusselt number Nuavg and pressure drop DelP_Pa from seven operating and geometry features, following the authors' version 4 data supplement for the associated paper.

Two independent 800-tree forests were fitted on 3,686 samples and tested on 922 samples. weight/best_model.pth is a joblib bundle, despite the .pth extension; it is not a PyTorch state dictionary. The published bundle changes only data-path strings in its metadata. Its fitted forests are unchanged from the original run. The published SHA256 is d43fe39fc6de27aa26eeb063a8b81c2089150e09a7ea21db1e67d8f136aeab17; the original training checkpoint SHA256 was dad800c2ee630455ac421b81c5fc38a0a584c065adca720d0c534d2df065e421.

Intended Use

Task	Description
Thermo-hydraulic surrogate	Predict heat-transfer and pressure-drop responses
Design screening	Compare configurations within the training data's parameter range

Usage

Download the Model Package

hf download OneScience-Group/WavyChannel-RF-ThermoHydraulics --local-dir ./WavyChannel-RF-ThermoHydraulics
cd WavyChannel-RF-ThermoHydraulics

Install Dependencies

pip install -r requirements.txt
pip install 'modelscope[datasets]'

Use scikit-learn 1.6.0, as pinned in requirements.txt, when loading the joblib checkpoint. Only load model files from a trusted source.

Data

Hugging Face hosts only this model package. The two CSV files required by the fixed split are available as OneScience/WavyChannel-ThermoHydraulics on ModelScope:

modelscope download --dataset OneScience/WavyChannel-ThermoHydraulics --local_dir ./data

The package expects data/02_processed_data/ML_dataset_longform.csv and data/03_splits/random80_20_split.csv. Keep the authors' case_uid join and fixed train/test labels; do not generate a new split. The source is Mendeley Data version 4, CC BY 4.0.

Train

On a SLURM compute cluster, from the package root:

sbatch scripts/run_experiment.slurm

Alternatively, on an allocated compute node:

python scripts/train.py --config config/config.yaml
python scripts/inference.py --config config/config.yaml
python scripts/result.py --config config/config.yaml

Preserve the bundled weight and results before retraining. A Random Forest has no epochs; this is one full fit of two independent targets.

Results

The reproduced test MAE/RMSE were 0.083817/0.276695 for Nuavg and 0.640271/2.000163 Pa for pressure drop. The data supplement reports 0.083819/0.276695 and 0.639853/1.997940 Pa, respectively. The archived validation JSON files refer to the original training-file hash; reloaded predictions from the published bundle must be checked separately.

OneScience Repositories

Platform	Main repository	Skills repository
Gitee	OneScience	OneSkills
GitHub	OneScience	OneSkills

Citation and License

• Paper: Machine learning surrogate for CFD-based prediction of heat transfer and pressure drop in partially porous wavy channels.
• Data: Kumar and Pandey, Mendeley Data v4, CC BY 4.0. Model-code and dataset licenses are separate.