UARK-NED3/ColdPlate-CFD-2000sims

Cold Plate CFD: 2000 Simulations

2,000 steady-state CFD simulations of a liquid-cooled, dual-chip, pin-fin cold plate, plus the POD-NN surrogate trained on them and the code that reproduces every figure in the paper.

Dataset DOI: 10.5061/dryad.k0p2ngfp5 · Primary article: AI Thermal Fluids, 10.1016/j.aitf.2026.100040 · Model: UARK-NED3/PODNN-ColdPlate · License: CC0 1.0 (public domain dedication)

Mirror of the Dryad deposit published 29 April 2026.

What was simulated

Thermal management sets the ceiling on power density in modern electronics, and CFD is accurate enough to design against but far too slow to sit inside an optimization loop or a live digital twin. This dataset was built to train and benchmark surrogates that close that speed gap.

The CFD model solves the steady, incompressible Navier–Stokes equations coupled with the energy equation using Menter's k-ω SST turbulence model on a roughly 50,000-element three-dimensional grid in ANSYS Fluent. The design of experiments varies six boundary conditions across the cold plate's operational space by Latin Hypercube Sampling:

Parameter	Boundary
chip1.thermal.heat_flux	chip 1
chip2.thermal.heat_flux	chip 2
inlet1.momentum.mass_flow_rate	inlet 1
inlet1.thermal.total_temperature	inlet 1
inlet2.momentum.mass_flow_rate	inlet 2
inlet2.thermal.total_temperature	inlet 2

The published surrogate uses only the central yz mid-plane (11,110 nodes). The other stored surfaces are included for reuse but are not exercised by the paper.

Headline results from the paper

• POD with 4 retained modes captures 99.32% of the temperature-field variance.
• The trained POD-NN predicts a full 2D temperature field 412,500× faster than the underlying CFD solve.
• A 100-sample training set is enough to reach the error floor; the iteration-cost crossover with CFD occurs at 100.5 design iterations (about 2.76 hours).

Contents

1_Dataset/

sim_0001.npz through sim_2000.npz — one NumPy archive per simulation. The i-th file corresponds to the i-th boundary-condition row in model_setup.json. Each holds 14 float-array keys:

Key	Shape	Meaning
yz-mid|temperature	(11110,)	Temperature (K) on the central yz mid-plane — the field used in the paper
yz-mid|coordinates	(11110, 3)	X/Y/Z node coordinates (m)
zx-mid|temperature	(40939,)	Temperature (K) on the zx mid-plane
zx-mid|coordinates	(40939, 3)	X/Y/Z node coordinates (m)
bottom|temperature	(3193,)	Temperature (K) on the bottom-face cooling boundary
bottom|coordinates	(3193, 3)	X/Y/Z node coordinates (m)
chip1_tavg|temperature, chip1_tmax|temperature	(1,)	Chip 1 mean / max temperature
chip2_tavg|temperature, chip2_tmax|temperature	(1,)	Chip 2 mean / max temperature
outlet_tavg|temperature, outlet_tmax|temperature	(1,)	Outlet mean / max temperature
pdrop_1|temperature, pdrop_2|temperature	(1,)	Pressure drop, circuits 1 and 2

Two notes on the scalar channels, both verified against the released files:

• The pdrop_* keys carry pressure drop, not temperature. The |temperature suffix is an artifact of the original export naming — every scalar report was written with it. Sampled values span roughly 4.2 to 12.0 while temperatures sit at 300–480 K. model_setup.json lists pdrop_1 and pdrop_2 as their own Fluent Report Definitions. The pressure unit is not recorded anywhere in the deposit, so treat these as relative unless you recover the unit from the original case setup.
• outlet_tavg|temperature and outlet_tmax|temperature are identical in every one of 120 randomly sampled simulations. Treat them as one channel; training on both double-weights the same quantity.

model_setup.json — the DOE configuration from the ANSYS/PyFluent workflow. Every script in 3_SourceCode/ reads it to recover per-sample input values. Top-level keys are timestamp, model_inputs, model_outputs, doe_configuration (nested as {boundary: {param: [value_1 … value_2000]}}), and case_file. The last is an informational path only — the original ANSYS .cas.h5 case file is not redistributed here.

2_AnalyzedData/

trained_model/ — the canonical POD-NN surrogate: pod_nn.h5 (Keras weights), pca.pkl (4 POD modes and the snapshot mean), param_scaler.pkl, mode_scaler.pkl, and training_history.json. Published on its own at UARK-NED3/PODNN-ColdPlate and kept here so this mirror stays complete against the deposit.

Trained on sim_0001–sim_0100, 4 modes, up to 500 epochs, seed 42, batch size 8, 80/20 train/validation, with sim_1601–sim_2000 held out. Test metrics: R² = 0.998, RMSE = 1.94 K, MAE = 1.46 K.

dataset_size_sensitivity.csv — the MSE / Max AE / MSE_POD / MSE_NN numbers behind the sensitivity and error-decomposition plots (script 09).

3_SourceCode/

_utils.py holds data loading, POD-NN build/train/predict, artifact save and load, thermal-resistance computation, and shared plotting style. All hyperparameters are module constants there: TRAIN_SIZE=100, TEST_SIZE=400, N_MODES=4, EPOCHS=500, RANDOM_SEED=42, CHIP_AREA=0.0016 m².

Script	Produces
01_train_model.py	Trains the canonical POD-NN and writes 2_AnalyzedData/trained_model/. Run first.
02_pod_modes.py	POD spatial mode plots
03_pod_reconstruction.py	Original / 1-PC / 4-PC reconstruction quad plot
04_pod_variance.py	Cumulative variance vs mode count
05_mse_history.py	Train / validation MSE curve
06_time_crossover.py	CFD-vs-surrogate time crossover and per-component bar chart
07_single_prediction.py	Per-sample prediction / truth / error plots
08_mean_error_field.py	MAE field on the yz mid-plane across the test set
09_dataset_size_sensitivity.py	Sensitivity sweep and error decomposition (slow — trains 10 models)
10_pod_alpha_sweep.py	POD mode α-sweep heatmaps and per-mode delta plot
11_pearson_correlation.py	POD coefficient vs physical-quantity Pearson R heatmap
12_rth_latent_space.py	Thermal-resistance response surfaces in PC3–PC4

Loading

The Dataset Viewer is disabled: the payload is .npz archives, which the viewer cannot render. Browse the Files tab, or pull the tree:

from huggingface_hub import snapshot_download
import numpy as np

path = snapshot_download("UARK-NED3/ColdPlate-CFD-2000sims", repo_type="dataset")

d = np.load(f"{path}/1_Dataset/sim_0001.npz", allow_pickle=True)
T   = d["yz-mid|temperature"]     # (11110,) K
xyz = d["yz-mid|coordinates"]     # (11110, 3) m
print(T.shape, T.min(), T.max())

To fetch only the simulations and DOE without the code or model:

path = snapshot_download(
    "UARK-NED3/ColdPlate-CFD-2000sims", repo_type="dataset",
    allow_patterns=["1_Dataset/*"],
)

.npz needs NumPy ≥ 1.20. The .pkl files unpickle into scikit-learn objects and therefore execute code on load — they are published unmodified from the CC0 deposit; load them only from here or Dryad. The .h5 file loads with tensorflow.keras.models.load_model(path, compile=False).

Reproducing the paper

pip install -r 3_SourceCode/requirements.txt
cd 3_SourceCode
python 01_train_model.py     # must run first

Then run scripts 02 through 12 in any order. Script 09 takes a few minutes; the rest finish in seconds. Every script seeds NumPy and TensorFlow with RANDOM_SEED=42, the first 100 simulations train, the last 400 test, and the 80/20 validation split is fixed — so with matching Python and TensorFlow versions, repeated runs reproduce the published numbers.

One caveat on that: requirements.txt lists bare package names with no version pins, and no lockfile or environment capture ships with the deposit. The matching versions the reproducibility note depends on are therefore not recoverable from the release itself. The one version that is recoverable is scikit-learn 1.7.2, embedded in the _sklearn_version field of the three .pkl artifacts.

Relationship to CFDTwin

CFDTwin (docs, Zenodo 10.5281/zenodo.20249626) is the production successor: a wizard-driven desktop GUI that automates this POD-NN pipeline end to end against a live ANSYS Fluent case, in five steps — Setup, DOE, Simulate, Train, Validate.

The software is described in Curl, D. and Hu, H. CFDTwin: An open-source GUI and Python toolkit for POD-NN surrogate modeling of ANSYS Fluent simulations. arXiv:2605.27725 (2026). arXiv · doi:10.48550/arXiv.2605.27725

These are not interchangeable. The sim_NNNN.npz + model_setup.json format here is a snapshot from an earlier iteration of the codebase and is not loadable by the current CFDTwin release. The scripts in 3_SourceCode/ are self-contained and reproduce the paper from this static dataset. For new CFD cases, run CFDTwin against your own Fluent setup rather than retrofitting this format. CFDTwin is also licensed differently — MIT, where this deposit is CC0.

Funding

U.S. National Science Foundation award OIA-2429580, EPSCoR Research Fellow: NSF: Immersion Cooling of Interior Permanent Magnet Synchronous Motors with Additively Manufactured Stator Windings (award page).

Note on the deposit README

The README published standalone on Dryad records the dataset DOI, while the copy bundled inside ned-009_AIEmulator.zip carries an unfilled Dataset DOI: FILL placeholder. This card uses the resolved DOI. The maintainers' working copy has since been corrected, for a future deposit version.

Citation

Curl, D. and Hu, H. 2026. Physically interpretable surrogate modeling of thermal fields in electronics cooling using combined proper orthogonal decomposition and neural networks. AI Thermal Fluids 6, 100040. https://doi.org/10.1016/j.aitf.2026.100040

Data: Curl, Daniel and Han Hu. Data from: Physically interpretable surrogate modeling of thermal fields in electronics cooling using combined proper orthogonal decomposition and neural networks. Dryad. https://doi.org/10.5061/dryad.k0p2ngfp5

Contact

Han Hu, Associate Professor of Mechanical Engineering, University of Arkansas — hanhu@uark.edu

Start here

What this resource supports. This release provides 2,000 steady-state CFD simulations of a liquid-cooled dual-chip pin-fin cold plate, the associated POD-NN surrogate artifacts, and the source code used to reproduce the paper figures. The published surrogate operates on the central yz mid-plane temperature field. First five minutes. Download the release tree, load one simulation archive with NumPy, and inspect model_setup.json for the corresponding six boundary-condition inputs. For the published static release, reproduce the analysis with 3_SourceCode/01_train_model.py before running the figure scripts. The companion model card is https://huggingface.co/UARK-NED3/PODNN-ColdPlate . Use with care. The simulations are steady-state, incompressible CFD under the documented modeling assumptions. The pdrop_* channels are pressure drops despite their exported suffix, and their unit is not recorded in the deposit. The current static format is not interchangeable with CFDTwin; for new Fluent cases, use CFDTwin with your own verified setup rather than retrofitting this archive. Continue. Dataset DOI: https://doi.org/10.5061/dryad.k0p2ngfp5 NED³ software catalog: https://ned3.uark.edu/software/