NurshatMenglik/NeuralConvection_3D_TrainTest

NeuralConvection — 3D Training/Test Data (version 2)

📄 Paper: A Neural Surrogate Approach for Simulating Natural Convection Problems (arXiv:2606.25259) — Nurshat Menglik, Alex Shao, David Hyde.

Version 2 (September 2026) replaces the original release. All pairs were regenerated; see What changed in version 2. The original tensors remain available unchanged under the tag v1.0-legacy (snapshot_download(..., revision="v1.0-legacy")). New work should use version 2.

Paired 3D Boussinesq → compressible steady natural-convection fields in the unit cube, used to train and evaluate the 3D Fourier Neural Operator (FNO) surrogate of the paper. Each pair solves the same randomized problem with both models on the same $24^3$ mesh.

Each split provides

• X.pt — Boussinesq input, shape (N, 5, 24, 24, 24), float32
• Y.pt — compressible target, same shape
• kept_samples.txt, per_sample.csv, MANIFEST.json — sample ids, per-sample parameters (Ra, Ma, seed, unit-conversion factors, steady-state record, input-vs-target differences) and the packing conventions

Channels of both tensors: [u_x, u_y, u_z, T, p] on a uniform $24^3$ grid spanning the cube; tensor axes are (x, y, z) and gravity points along -y.

Splits

Folder	N	Description
train/	1,000	Stochastic Voronoi wall-temperature generator, 2–5 Voronoi cells per active wall
test_in_distribution/	50	Same generator, independent draws
test_ood/	50	6–9 Voronoi cells per active wall: wall patterns richer than any seen in training, same Rayleigh range
test_legacy/	50	Constant hot and cold side walls, the other four walls adiabatic

Every split draws Ra in [1e2, 1e5], the range in which the $24^3$ discretization reproduces reference Nusselt numbers to within one percent; Pr = 0.71, wall temperatures in [0.4, 1.6] (T_init = 1, ΔT = 1.2).

train/split_train_val.json gives the 900/100 training/validation split used for the models in the paper. train/flagged_samples.json lists 24 training pairs, all at Ra ≤ 222, whose residual solver drift is comparable to the small Boussinesq–compressible difference they carry; they were used in the paper and are kept here, so filter them if you prefer.

Units and conventions

• Both members are in one unit system, the buoyancy (free-fall) scales used by the compressible model: velocity in U = sqrt(gβΔT·L), pressure in ρ₀U², temperature as T/T₀. The Boussinesq model is non-dimensionalized with this same velocity scale (its coefficients are sqrt(Pr/Ra), 1 and 1/sqrt(Ra·Pr) instead of the Pr, Pr·Ra, 1 of the thermal-diffusion convention), so no unit change is applied to either member and the pair differs by physics only. To recover the thermal-diffusion (de Vahl Davis) convention multiply velocities by sqrt(Ra·Pr) and pressures by Ra·Pr (vel_factor in per_sample.csv).
• Pressure is the solver's pressure perturbation with its spatial mean and its linear-in-y hydrostatic component removed, for both members; the Boussinesq pressure is stored with the sign convention of the compressible one.
• Values are not normalized. The models in the paper use a per-sample, per-channel affine map taken from the min–max range of the sample's own input X and applied identically to X and Y, and they predict a residual, y = x + f(x).
• Both members are marched to a steady state (per-step relative change below 1e-6); per_sample.csv records the achieved change of each member.

What changed in version 2

1. Boussinesq buoyancy term. Version 1 assembled the Boussinesq body force with T/T_init instead of (T − T_init)/ΔT; with ΔT = 1.2 that is an effective Rayleigh number of 1.2·Ra for the Boussinesq member while the compressible member ran at Ra. Version 2 uses (T − T_init)/ΔT.
2. Steady states replace the fixed end time of version 1.
3. One unit system and a physical pressure channel (see above). Version 1 stored the Boussinesq input in thermal-diffusion units and the compressible target in free-fall units, with the pressure min–max normalized to [0, 1].
4. Rayleigh range restricted to [1e2, 1e5] (see above); the out-of-distribution axis is now the complexity of the wall pattern, not the Rayleigh number.
5. Split sizes are 1,000 / 50 / 50 / 50 (version 1: 1,005 / 106 / 106 / 120).

Loading

import torch
X, Y = torch.load("train/X.pt"), torch.load("train/Y.pt")           # (1000, 5, 24, 24, 24)

# normalization used in the paper: per-sample, per-channel min-max of the INPUT, applied to both
mn = X.amin(dim=(2, 3, 4), keepdim=True)
rg = (X.amax(dim=(2, 3, 4), keepdim=True) - mn).clamp_min(1e-12)
Xn, Yn = (X - mn) / rg, (Y - mn) / rg

The raw finite-element output of the 2D pipeline at four resolutions is in Paired_Boussinesq_Compressible_Dataset; the 2D tensors are in NeuralConvection_2D_TrainTest.

Source

Generated with the solver and data-generation pipeline at https://github.com/Nurshat317/NeuralConvection

License

Released under CC BY 4.0. You are free to share and adapt this data, including commercially, provided you give appropriate credit and indicate any changes. The requested form of credit is a citation to the paper below.

Citation

If you use this dataset in your research, please cite:

@article{menglik2026neural,
  title         = {A Neural Surrogate Approach for Simulating Natural Convection Problems},
  author        = {Menglik, Nurshat and Shao, Alex and Hyde, David},
  journal       = {arXiv preprint arXiv:2606.25259},
  year          = {2026},
  eprint        = {2606.25259},
  archivePrefix = {arXiv},
  primaryClass  = {physics.comp-ph},
  url           = {https://arxiv.org/abs/2606.25259}
}

• Paper: arXiv:2606.25259
• Code: github.com/Nurshat317/NeuralConvection