acceptodds
Under review as a conference paper at ICLR 2027

CoolData: Benchmarking Machine Learning Methods for Electronics Cooling

Abstract

The integration of Machine Learning (ML) into Computer-Aided Engineering (CAE) workflows promises significant speed-ups, particularly in the field of Computational Fluid Dynamics (CFD). However, the development of robust ML models in CFD is hindered by the lack of high-quality, application-specific datasets. This work addresses this gap with CoolData, a large-scale electronics cooling dataset containing stationary 3D flow and temperature fields for a diverse set of geometries, simulated using the industrial CFD solver Simcenter STAR-CCM+. To our knowledge, this is the first high-quality and publicly available dataset for electronics cooling applications. We show how this dataset can be leveraged in an ML pipeline, benchmarking both volumetric models (3D U-Net, DOMINO, Transolver) and surface models (MeshGraphNet, DOMINO, Transolver) across a range of training-set sizes. Beyond in-distribution accuracy, we introduce a systematic out-of-distribution (OOD) study along two axes enabled by CoolData's controlled parametrization: geometric composition (body count and shape held out from training) and thermal extrapolation (peak component temperature beyond the training range).

Primary area:
datasets and benchmarks
Keywords:
electronics cooling, CFD, dataset, machine learning, benchmarking

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CCc2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces an electronics dataset, CoolData. It contains 60848 steady 3D CFD solutions for electronics-cooling-like flows, computed with the commercial CFD software Simcenter STAR-CCM+. The simulation setup is a 0.5x0.1x0.02m air channel with 2 to 6 heated cuboid or cylinder blocks mounted on the bottom wall, solved under steady RANS with k-epsilon turbulent model. The meshs are trimmed hex cells with a 2mm base size and about 160k cells per case. Four baselines were trained and compared mainly in two fields: Temperature and surface wall shear stress.

Critical strengths
  1. The dataset stores quite rich outputs, such as HTC, wall shear stress, pressure, and temperature.
Critical weaknesses
  1. The simulated solutions of the CoolData are neither validated nor carried out mesh independence study, both are standard practice in CFD. The only validation cited is an ORNL/CASL report on LWR (nuclear) cases (ref 30). It is unrelated to this setup, and validation of a solver does not automatically ensure a physically correct solution because other factors such as mesh resolution, initial and boundary conditions, and the spatial and temporal discretization schemes will significantly affect the solution, especially RANS is considered.
  2. "high-fidelity" is overstated by the authors. Considering this is a 3D dataset, the mesh is way too coarse relative to the geometry, and the mesh size of each case is just 160k cells, a lot of 2D cases can easily make up of at least 300k cells. Thus, the comparison to 100GB car simulation is rhetorical. In addition, the setup is also very idealised in which the authors considered no conduction or radiation (although these two are also the main heat transfer paths besides convection), fixed-T blocks, no PCB or housing, and no fans.
  3. There is no engineering-relavant metrics. The paper stresses Q_total, thermal resistance and pressure drop as the point of the dataset, but never reports surrogate errors on them. Errors are pooled over all points, so the uniform inlet region dominates.
  4. The presentation of figures are poorly done. Hard to read and wrong information. For example, Figure 10 shows duplicated HTC maps (the same mean/std grids as Figures 14 and 7). Its caption describes cumulative distributions that actually appear in Figure 11.
Questions for authors
  1. What validation evidence exists for this exact setup or comparison with a finer mesh, a different turbulence model, or LES on a few cases?
  2. Why benchmark wall shear stress rather than HTC or heat flux since this work's scope is electronic cooling?
  3. Can you report engineering/physics metrics, like error in Q_total, max component temperature, pressure drop, area-averaged HTC, and error in the near-body region rather than only pooled global MAE?
Reviewer_ZcyVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CoolData, a large-scale CFD dataset for electronics cooling, comprising steady-state 3D flow and temperature simulations with diverse geometric configurations. The dataset is generated using the industrial CFD solver Simcenter STAR-CCM+ and supports systematic evaluations of data efficiency and out-of-distribution generalization. The authors benchmark several machine learning models on volumetric temperature prediction and surface wall shear stress prediction.

Critical strengths
  1. Large-scale and valuable dataset. The paper introduces a large-scale dataset containing more than 60,000 three-dimensional CFD simulations, providing a valuable resource for applying machine learning to electronics cooling problems. Given the high computational cost of generating CFD simulation data, the scale of this dataset is a meaningful contribution.

  2. Practical relevance to industrial applications. The paper focuses on electronics cooling, an important engineering problem closely related to practical industrial applications.

Critical weaknesses
  1. Limited benchmark coverage and insufficient experimental evaluation. The experimental evaluation is relatively limited for a benchmark paper of this scale. Despite introducing a large dataset containing more than 60,000 CFD simulations with multiple physical variables, the paper primarily focuses on only two prediction tasks: volumetric temperature prediction and surface wall shear stress prediction. Moreover, each task includes only three baseline models, which is insufficient to comprehensively evaluate the capabilities of different machine learning approaches. In particular, several important physical quantities, such as velocity, pressure, and heat transfer coefficients, are available in the dataset but are not included in the main benchmarks.

  2. Concerns about the evaluation metrics and conclusions drawn from OOD experiments. The OOD experiments primarily rely on absolute errors (MAE) to compare prediction accuracy across different geometric configurations and thermal conditions. Although relative errors are provided in the appendix, the temperature prediction errors are normalized using a single global reference value rather than the characteristic signal magnitude of each OOD scenario. Since different configurations may exhibit substantially different temperature variations and wall shear stress magnitudes, comparisons based primarily on absolute errors or globally normalized errors may not adequately reflect the relative prediction difficulty of each scenario. Consequently, some conclusions about the relationship between geometric complexity, OOD shifts, and prediction difficulty are not fully convincing. I suggest reporting scenario-specific normalized errors, such as relative L2 errors or errors normalized by the characteristic temperature rise within each test group, and revisiting the conclusions based on these additional metrics.

  3. Poor readability of several figures. Many figures contain text, axis labels, legends, and numerical annotations that are too small to read comfortably, particularly those presenting multiple experimental results or learning curves. This makes it difficult to interpret the visualizations and compare results across different experimental settings.

  4. Potential selection bias introduced by discarding non-converged simulations. The paper mentions that CFD simulations failing to converge within 1,500 iterations are excluded from the dataset. However, these non-converged cases may correspond to particularly challenging geometric configurations or complex flow phenomena, which could be especially valuable for evaluating the robustness and generalization capabilities of machine learning models. Excluding such cases may introduce selection bias, making the resulting dataset disproportionately representative of configurations that are easier for the numerical solver to handle and potentially reducing the difficulty and practical value of the benchmark. It would be helpful if the authors could report the number and characteristics of the discarded samples and analyze how their exclusion affects the overall data distribution. Have the authors attempted to recover these cases through mesh refinement, alternative solver settings, or more robust numerical schemes? Additionally, have they considered the potential risks of systematically excluding difficult cases, particularly when the benchmark is intended to support real-world engineering applications?

Questions for authors

See above.

Reviewer_oPFWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This is a dataset and benchmark paper. The paper releases CoolData, a dataset of 60,848 steady 3D RANS (k–ε) simulations of forced-air cooling of electronics, made with STAR-CCM+. Each sample is a fixed channel with 2 to 6 heated blocks or cylinders attached to the wall. Their positions, sizes and surface temperatures vary, and so does the inlet speed (up to 35 parameters, Latin Hypercube sampling). Two benchmarks: predicting the temperature field in the volume (3D U-Net with FiLM, Transolver, DOMINO) and predicting wall shear stress on the surface (MeshGraphNet, Transolver, DOMINO), with 1k to 50k training samples. Two out-of-distribution studies: body composition (train on 3 blocks only, test on other combinations of body count and shape) and temperature extrapolation (train on data where the hottest body is below 60 °C).

Critical strengths
  1. Dataset: a large, open dataset in an under-served area, with a concrete and reproducible generation pipeline. Existing industrial-solver CFD datasets are almost all external aerodynamics. Internal forced convection over heated components is a gap. The release includes 60k samples, the solver setup and an accessor library. The convergence rule is spelled out with numbers: the mean velocity, mean temperature and outlet heat transfer must each stay within 0.0001, 0.00001 and 0.0001 over the last 30 iterations, there is a residual stability check, and cases that don't converge or oscillate are removed.

  2. Benchmark: a controlled parametrization that allows sensible data-scaling and out-of-distribution studies.The nested training subsets share fixed validation and test sets, and the out-of-distribution splits follow interpretable axes (body count and shape, temperature, and others in Fig. 11). That makes the benchmark more useful than an unstructured pile of simulations.

Critical weaknesses

Dataset:

  1. There is no explanation regarding data reliability, the claim of high fidelity lacks verification at the data level, for instance, in terms of adherence to physical properties or statistical characteristics.

  2. The benchmark conclusions are not sufficiently reliable. It appears that only a single random seed was used, the "± std" value represents the dispersion across test samples, making it impossible to distinguish between models like DOMINO and Transolver when their performance differs by only 0.1 K. Furthermore, looks like 3D U-Net was evaluated using a voxelized target and other models are evaluated at the centers of the grid cells.

Questions for authors

How many physical properties of the data can your STAR-CCM+ account for?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.