{"success":true,"data":{"author":"Nurshat Menglik, Alex Shao, David Hyde","bucketAccessType":"public","bucketName":"NurshatMenglik/NeuralConvection_2D_TrainTest","bucketPrefix":"","category":"Physics","cloudProvider":"huggingface","createdAt":"2026-11-21T20:46:01.200Z","description":"Training and test tensors for the 2D Fourier Neural Operator (FNO) surrogate in \"A Neural Surrogate Approach for Simulating Natural Convection Problems\" (Menglik, Shao and Hyde, arXiv:2606.25259). Each pair maps a steady Boussinesq natural-convection field to the fully compressible field for the same randomized problem in the unit-square cavity, both solved with FEniCSx on one 64 x 64 mesh. X.pt (Boussinesq input) and Y.pt (compressible target) are float32 PyTorch tensors of shape (N, 4, 64, 64) with channels u_x, u_y, temperature T and pressure p; gravity points along -y and column 0 is the hot wall. Splits: train (4,000 stochastic Fourier wall-temperature profiles), train_legacy (5,000 classical hot-left/cold-right cavities), and 200-sample in-distribution, out-of-distribution (richer wall patterns), shifted-amplitude and legacy test sets, with disjoint seed streams. Rayleigh number Ra is log-uniform from 1e2 to 3e6 and Prandtl number 0.71; per_sample.csv, kept_samples.txt and MANIFEST.json give per-sample parameters and conventions. Version 2 (September 2026), legacy tensors under tag v1.0-legacy. Load with torch.load. Released under CC BY 4.0; cite arXiv:2606.25259.","displayTitle":"NeuralConvection 2D: Boussinesq-to-Compressible Natural Convection Tensors for Fourier Neural Operators (Menglik, Shao, Hyde)","huggingFaceUrl":"https://huggingface.co/datasets/NurshatMenglik/NeuralConvection_2D_TrainTest","id":"9858eb61-06d4-5e4b-b32b-7b28bef6fc37","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/NurshatMenglik/NeuralConvection_2D_TrainTest","objectCountEstimate":33,"probeStatus":"ok","referenceUrl":"https://arxiv.org/abs/2606.25259","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"NurshatMenglik/NeuralConvection_2D_TrainTest\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"NurshatMenglik/NeuralConvection_2D_TrainTest\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=NurshatMenglik, license:cc-by-4.0, size_categories:1K<n<10K, downloads=181"],"format":"mixed","generatedAt":"2026-11-21T20:48:51.579Z","layout":{"example":"test_in_distribution/per_sample.csv","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"sample_id,legacy_row,row,Ra,log10Ra,Pr,Fr,gamma,eps,T0,dT,sqrt_RaPr,vel_factor,a_B,b_B,s_B,m_in,a_c,b_c,s_c,k_c,m_tgt,mean_T_B,mean_invT_c,rms_r_B,rms_r_c,ra_method,ra_confidence,umax_in,umax_tgt,ratio_u,T_min_in,T_max_in,T_min_tgt,T_max_tgt,p_rms_in,p_rms_tgt,p_corr_in_tgt,relL2_u_x,relL2_u_y,relL2_T,relL2_p\r\nsample_000000,0,0,181738.28211307156,5.259446418553411,0.71,1.095,1.4,0.6,1.0,1.2,359.2132796825318,1.0,-0.002521593369443714,0.005007061637034089,1.0,1.8973538018496328e-19,0.4142617484188994,-0.8284","encoding":"utf8","key":"test_in_distribution/per_sample.csv","truncatedAt":512}],"sampleObjects":["test_in_distribution/X.pt","test_in_distribution/Y.pt","test_in_distribution/kept_samples.txt","test_in_distribution/per_sample.csv","test_legacy/X.pt"],"sidecars":[{"bytes":"---\nlicense: cc-by-4.0\npretty_name: NeuralConvection 2D Train/Test\nsize_categories:\n- 1K<n<10K\ntags:\n- computational-fluid-dynamics\n- natural-convection\n- boussinesq\n- compressible-flow\n- fourier-neural-operator\n- scientific-machine-learning\n---\n\n# NeuralConvection — 2D Training/Test Data (version 2)\n\n📄 **Paper:** [A Neural Surrogate Approach for Simulating Natural Convection\nProblems (arXiv:2606.25259)](https://arxiv.org/abs/2606.25259) — Nurshat Menglik,\nAlex Shao, David Hyde.\n\n> **Version 2 (September 2026) replaces the original release.** All pairs were regenerated;\n> see [What changed in version 2](#what-changed-in-version-2). The original tensors remain\n> available unchanged under the tag `v1.0-legacy`\n> (`snapshot_download(..., revision=\"v1.0-legacy\")`). New work should use version 2.\n\nPaired 2D Boussinesq → compressible steady natural-convection fields in the unit square, used to\ntrain and evaluate the 2D Fourier Neural Operator (FNO) surrogate of the paper. Each pair solves\nthe same randomized problem with both models on the same 64 × 64 mesh.\n\nEach split provides\n\n- `X.pt` — Boussinesq input, shape `(N, 4, 64, 64)`, float32\n- `Y.pt` — compressible target, same shape\n- `kept_samples.txt`, `per_sample.csv`, `MANIFEST.json` — sample ids, per-sample parameters\n  (`Ra`, `Ma`, wall dials, pressure de-ramping coefficients, input-vs-target differences) and\n  the packing conventions\n\nChannels of both tensors: `[u_x, u_y, T, p]` on a uniform 64 × 64 grid; row index = `y`\n(row 0 at the bottom), column index = `x` (column 0 at the hot wall); gravity points along `−y`.\n\n## Splits\n\n| Folder | N | Description |\n|---|---|---|\n| `train/` | 4,000 | Stochastic Fourier wall temperatures (per-wall complexity dials; adiabatic top/bottom walls possible) |\n| `train_legacy/` | 5,000 | Constant hot left / cold right wall, adiabatic top and bottom (the classical cavity) |\n| `test_in_distribution/` | 200 | Same generator as `train/`, independent seed stream |\n| `test_ood/` | 200 | Same generator with the vertical-wall dial drawn from {4,5,6} (3–5 Fourier modes per side wall) against {1,2,3} in `train/`; horizontal walls, amplitude band and `Ra` law as in `train/`. The out-of-distribution set of the paper (Test Set 3); no constant-wall samples can occur |\n| `test_rand_shift/` | 200 | Same generator with the admissible wall-amplitude band halved: a covariate shift inside the training support. Kept for the record; no longer used in the paper, which uses `test_ood/` as its third 2D test set |\n| `test_legacy/` | 200 | Classical cavity, independent seed stream |\n\n`Ra` is log-uniform in `[1e2, 3e6]` (capped at `2e6` when a horizontal wall is conducting),\n`Pr = 0.71`, wall temperatures in `[0.4, 1.6]` (`T_init = 1`, `ΔT = 1.2`). The six seed streams are\ndisjoint, so no problem appears in two splits. `train/flagged_samples.json` lists two training pairs\nwhose Rayleigh number was lowered because the canonical draw did not reach a steady state.\n\nNote: `test_in_distribution/` shares its seed stream with the res-32 block of the\n[10,000-sample release](https://huggingface.co/datasets/NurshatMenglik/Paired_Boussinesq_Compressible_Dataset)\n(its first 200 samples are the same problems at 32 × 32); do not train on that block if you evaluate\non this split.\n\n## Units and conventions\n\n- **Both members are in one unit system**, the buoyancy (free-fall) scales used by the\n  compressible model: velocity in `U = sqrt(gβΔT·L)`, pressure in `ρ₀U²`, temperature as `T/T₀`.\n  The Boussinesq model is non-dimensionalized with this same velocity scale (its coefficients are\n  `sqrt(Pr/Ra)`, `1` and `1/sqrt(Ra·Pr)` instead of the `Pr`, `Pr·Ra`, `1` of the thermal-diffusion\n  convention), so no unit change is applied to either member and the pair differs by physics only.\n  To recover the thermal-diffusion (de Vahl Davis) convention multiply velocities by `sqrt(Ra·Pr)`\n  and pressures by `Ra·Pr` (`sqrt_RaPr` in `per_sample.csv`).\n- **Pressure** is the solver's raw pressure with its spatial mean and its linear-in-`y`\n  hydrostatic component","encoding":"utf8","key":"README.md","name":"README.md","size":7372,"truncated":true},{"bytes":"{\n \"manifest_version\": \"c3-phase1-v1\",\n \"created_utc\": \"2026-09-22T00:56:00+00:00\",\n \"git_commit\": \"v6_freefall_20260921\",\n \"split\": \"test_in_distribution\",\n \"units\": \"consistent\",\n \"common_system\": \"freefall\",\n \"pressure_mode\": \"raw_both_deramp\",\n \"boussinesq_pressure_origin\": \"raw solver field P_bouss_raw.pt (no ruler)\",\n \"compressible_pressure_origin\": \"raw solver field P_comp_raw.pt (no ruler, k_c = Ra*Pr exactly)\",\n \"deramped\": {\n  \"input\": true,\n  \"target\": true\n },\n \"ra_split_key\": \"test_rand\",\n \"tensor_layout\": \"(N, 4, ny, nx) float32; channels [u_x, u_y, T, p]; row index = y (row 0 bottom), column index = x (col 0 hot wall); y_j = j/(ny-1)\",\n \"shape\": [\n  200,\n  4,\n  64,\n  64\n ],\n \"dtype\": \"torch.float32\",\n \"constants\": {\n  \"Pr\": 0.71,\n  \"Fr\": 1.095,\n  \"gamma\": 1.4,\n  \"T0\": 1.0,\n  \"dT\": 1.2,\n  \"eps\": 0.6,\n  \"ref_slope_B\": \"-Pr*Ra\",\n  \"ref_slope_c\": -0.8340109672442193\n },\n \"formulas\": {\n  \"common_units\": \"free-fall: velocity U = sqrt(g*beta*dT*L), pressure rho0*U^2, temperature T/T0 -- BOTH members solved in these scales (Boussinesq member: genData_v6, coefficients sqrt(Pr/Ra), 1, 1/sqrt(Ra*Pr)); no unit change applied\",\n  \"velocity\": \"identity for both (u_in = u_B in free-fall units, u_tgt = u_c); vel_factor = 1\",\n  \"temperature\": \"identity for both (stored as T/T0 in [0.4, 1.6], eps = 0.6); range asserted\",\n  \"pressure_deramp\": \"for each field fit p01(x,y) ~ a + b*y by least squares over the 64x64 field, y = row/(ny-1); r = p01 - (a + b*y) (zero-mean residual)\",\n  \"pressure_sign\": \"the generator's Boussinesq weak form uses '+ p*div(v)', the opposite of the paper's, so the raw Boussinesq pressure is the negative of the paper's P: P_in = -(p_raw_B - mean). The compressible member keeps the paper sign: P_tgt = +Ra*Pr*(p_raw_C - mean).\",\n  \"pressure_center\": \"spatial mean subtracted from P_in and P_tgt after scaling (m_in, m_tgt recorded)\",\n  \"approximation\": \"raw_both: none on the scale (both members exact); the pair is asymmetric because only the compressible member carries a hydrostatic ramp.  raw_both_deramp: the linear-in-y fit is applied identically to both, which is symmetric but also removes physical linear-in-y dynamic-pressure content from each member.\",\n  \"pressure_scale\": \"INPUT: the RAW Boussinesq solver pressure (P_bouss_raw.pt, units rho0*alpha^2/L^2) is used directly, s_B = 1, no ruler -- the corrected buoyancy Pr*Ra*(T-T0)/dT leaves no hydrostatic ramp.  TARGET: the RAW compressible solver pressure (P_comp_raw.pt, units rho0*U_c^2) is used directly too, s_c = 1 -- NO RULER ON EITHER SIDE.  The only operation on the target is the unit change k_c = Ra*Pr, which is exact rather than the per-sample (1/Fr^2)/|b_c| estimate the ramp ruler had to make from a min-max-normalised field. Both members are de-ramped with the same least-squares linear-in-y fit, which removes the compressible hydrostatic stratification and the (physical) linear-in-y part of both dynamic pressures.  This is the like-for-like variant, directly comparable with the v2 ramp_ruler and v3 raw_bouss_deramp datasets.\"\n },\n \"conversion_options\": {\n  \"min_abs_slope\": 0.1,\n  \"chunk\": 500\n },\n \"counts\": {\n  \"n_input\": 200,\n  \"n_kept\": 200,\n  \"n_dropped\": 0\n },\n \"dropped_samples_missing_Ra\": [],\n \"positive_slope_samples\": [],\n \"temperature_range_check\": {\n  \"input\": [\n   0.4000000059604645,\n   1.600000023841858\n  ],\n  \"target\": [\n   0.4000000059604645,\n   1.600000023841858\n  ],\n  \"expected\": [\n   0.4,\n   1.6\n  ],\n  \"ok\": true\n },\n \"summary\": {\n  \"Ra_min\": 100.49267286819563,\n  \"Ra_max\": 2957309.692564718,\n  \"ratio_u_mean\": 1.0711074612076688,\n  \"ratio_u_min\": 0.7686269013559178,\n  \"ratio_u_max\": 1.3582580981027554,\n  \"p_corr_in_tgt_mean\": 0.9835481467857826,\n  \"p_corr_in_tgt_min\": 0.8512513725631204,\n  \"b_B_mean\": -0.0015820527474608747,\n  \"b_c_mean\": -0.8230924153064367,\n  \"b_B_is_slope_of\": \"raw Boussinesq pressure (expected ~0: no hydrostatic ramp)\",\n  \"b_B_over_PrRa_absmax\": 0.0007561839600364479,\n  \"b_c_is_slope_of\": \"raw compressible pressure (expected ~ -1/Fr^2 = -0.8337)\",\n  \"ruler_factor_c_if_it_had_been_used\": {\n   \"mean\": 1.0134345","encoding":"utf8","key":"test_in_distribution/MANIFEST.json","name":"MANIFEST.json","size":319656,"truncated":true},{"bytes":"{\n \"manifest_version\": \"c3-phase1-v1\",\n \"created_utc\": \"2026-09-22T00:56:00+00:00\",\n \"git_commit\": \"v6_freefall_20260921\",\n \"split\": \"test_legacy\",\n \"units\": \"consistent\",\n \"common_system\": \"freefall\",\n \"pressure_mode\": \"raw_both_deramp\",\n \"boussinesq_pressure_origin\": \"raw solver field P_bouss_raw.pt (no ruler)\",\n \"compressible_pressure_origin\": \"raw solver field P_comp_raw.pt (no ruler, k_c = Ra*Pr exactly)\",\n \"deramped\": {\n  \"input\": true,\n  \"target\": true\n },\n \"ra_split_key\": \"test_old\",\n \"tensor_layout\": \"(N, 4, ny, nx) float32; channels [u_x, u_y, T, p]; row index = y (row 0 bottom), column index = x (col 0 hot wall); y_j = j/(ny-1)\",\n \"shape\": [\n  200,\n  4,\n  64,\n  64\n ],\n \"dtype\": \"torch.float32\",\n \"constants\": {\n  \"Pr\": 0.71,\n  \"Fr\": 1.095,\n  \"gamma\": 1.4,\n  \"T0\": 1.0,\n  \"dT\": 1.2,\n  \"eps\": 0.6,\n  \"ref_slope_B\": \"-Pr*Ra\",\n  \"ref_slope_c\": -0.8340109672442193\n },\n \"formulas\": {\n  \"common_units\": \"free-fall: velocity U = sqrt(g*beta*dT*L), pressure rho0*U^2, temperature T/T0 -- BOTH members solved in these scales (Boussinesq member: genData_v6, coefficients sqrt(Pr/Ra), 1, 1/sqrt(Ra*Pr)); no unit change applied\",\n  \"velocity\": \"identity for both (u_in = u_B in free-fall units, u_tgt = u_c); vel_factor = 1\",\n  \"temperature\": \"identity for both (stored as T/T0 in [0.4, 1.6], eps = 0.6); range asserted\",\n  \"pressure_deramp\": \"for each field fit p01(x,y) ~ a + b*y by least squares over the 64x64 field, y = row/(ny-1); r = p01 - (a + b*y) (zero-mean residual)\",\n  \"pressure_sign\": \"the generator's Boussinesq weak form uses '+ p*div(v)', the opposite of the paper's, so the raw Boussinesq pressure is the negative of the paper's P: P_in = -(p_raw_B - mean). The compressible member keeps the paper sign: P_tgt = +Ra*Pr*(p_raw_C - mean).\",\n  \"pressure_center\": \"spatial mean subtracted from P_in and P_tgt after scaling (m_in, m_tgt recorded)\",\n  \"approximation\": \"raw_both: none on the scale (both members exact); the pair is asymmetric because only the compressible member carries a hydrostatic ramp.  raw_both_deramp: the linear-in-y fit is applied identically to both, which is symmetric but also removes physical linear-in-y dynamic-pressure content from each member.\",\n  \"pressure_scale\": \"INPUT: the RAW Boussinesq solver pressure (P_bouss_raw.pt, units rho0*alpha^2/L^2) is used directly, s_B = 1, no ruler -- the corrected buoyancy Pr*Ra*(T-T0)/dT leaves no hydrostatic ramp.  TARGET: the RAW compressible solver pressure (P_comp_raw.pt, units rho0*U_c^2) is used directly too, s_c = 1 -- NO RULER ON EITHER SIDE.  The only operation on the target is the unit change k_c = Ra*Pr, which is exact rather than the per-sample (1/Fr^2)/|b_c| estimate the ramp ruler had to make from a min-max-normalised field. Both members are de-ramped with the same least-squares linear-in-y fit, which removes the compressible hydrostatic stratification and the (physical) linear-in-y part of both dynamic pressures.  This is the like-for-like variant, directly comparable with the v2 ramp_ruler and v3 raw_bouss_deramp datasets.\"\n },\n \"conversion_options\": {\n  \"min_abs_slope\": 0.1,\n  \"chunk\": 500\n },\n \"counts\": {\n  \"n_input\": 200,\n  \"n_kept\": 200,\n  \"n_dropped\": 0\n },\n \"dropped_samples_missing_Ra\": [],\n \"positive_slope_samples\": [],\n \"temperature_range_check\": {\n  \"input\": [\n   0.4000000059604645,\n   1.600000023841858\n  ],\n  \"target\": [\n   0.4000000059604645,\n   1.600000023841858\n  ],\n  \"expected\": [\n   0.4,\n   1.6\n  ],\n  \"ok\": true\n },\n \"summary\": {\n  \"Ra_min\": 106.55828461551418,\n  \"Ra_max\": 2720347.173699123,\n  \"ratio_u_mean\": 1.1592755956565859,\n  \"ratio_u_min\": 1.1188248078257155,\n  \"ratio_u_max\": 1.203966212902729,\n  \"p_corr_in_tgt_mean\": 0.9817133134714262,\n  \"p_corr_in_tgt_min\": 0.9718658985291692,\n  \"b_B_mean\": 5.578809874999536e-15,\n  \"b_c_mean\": -0.8065267189601542,\n  \"b_B_is_slope_of\": \"raw Boussinesq pressure (expected ~0: no hydrostatic ramp)\",\n  \"b_B_over_PrRa_absmax\": 3.872131410417625e-17,\n  \"b_c_is_slope_of\": \"raw compressible pressure (expected ~ -1/Fr^2 = -0.8337)\",\n  \"ruler_factor_c_if_it_had_been_used\": {\n   \"mean\": 1.034109141817107,\n  ","encoding":"utf8","key":"test_legacy/MANIFEST.json","name":"MANIFEST.json","size":319121,"truncated":true},{"bytes":"{\n \"manifest_version\": \"c3-phase1-v1\",\n \"created_utc\": \"2026-09-25T10:09:49+00:00\",\n \"git_commit\": \"v6_freefall_ood_20260925\",\n \"code\": {\n  \"units.py\": \"/home/nurshat/FEniCS/c3_workspace/c3/units.py\",\n  \"convert_dataset.py\": \"/home/nurshat/FEniCS/c3_workspace/c3/convert_dataset.py\"\n },\n \"split\": \"test_ood\",\n \"units\": \"consistent\",\n \"common_system\": \"freefall\",\n \"pressure_mode\": \"raw_both_deramp\",\n \"source\": {\n  \"in_dir\": \"/data/nurshat/c3_workspace/data/v5_random/ff_tensors_ood/test_rand_ood\",\n  \"X\": \"/data/nurshat/c3_workspace/data/v5_random/ff_tensors_ood/test_rand_ood/X.pt\",\n  \"Y\": \"/data/nurshat/c3_workspace/data/v5_random/ff_tensors_ood/test_rand_ood/Y.pt\",\n  \"kept_samples\": \"/data/nurshat/c3_workspace/data/v5_random/ff_tensors_ood/test_rand_ood/kept_samples.txt\"\n },\n \"p_raw_source\": \"/data/nurshat/c3_workspace/data/v5_random/ff_tensors_ood/test_rand_ood/P_bouss_pert.pt\",\n \"p_raw_comp_source\": \"/data/nurshat/c3_workspace/data/v5_random/ff_tensors_ood/test_rand_ood/P_comp_pert.pt\",\n \"boussinesq_pressure_origin\": \"raw solver field P_bouss_raw.pt (no ruler)\",\n \"compressible_pressure_origin\": \"raw solver field P_comp_raw.pt (no ruler, k_c = Ra*Pr exactly)\",\n \"deramped\": {\n  \"input\": true,\n  \"target\": true\n },\n \"ra_json\": \"/data/nurshat/c3_workspace/data/v5_random/ff_tensors_ood/ra_v5.json\",\n \"ra_split_key\": \"test_rand_ood\",\n \"tensor_layout\": \"(N, 4, ny, nx) float32; channels [u_x, u_y, T, p]; row index = y (row 0 bottom), column index = x (col 0 hot wall); y_j = j/(ny-1)\",\n \"shape\": [\n  200,\n  4,\n  64,\n  64\n ],\n \"dtype\": \"torch.float32\",\n \"constants\": {\n  \"Pr\": 0.71,\n  \"Fr\": 1.095,\n  \"gamma\": 1.4,\n  \"T0\": 1.0,\n  \"dT\": 1.2,\n  \"eps\": 0.6,\n  \"ref_slope_B\": \"-Pr*Ra\",\n  \"ref_slope_c\": -0.8340109672442193\n },\n \"formulas\": {\n  \"common_units\": \"free-fall: velocity U = sqrt(g*beta*dT*L), pressure rho0*U^2, temperature T/T0 -- BOTH members solved in these scales (Boussinesq member: genData_v6, coefficients sqrt(Pr/Ra), 1, 1/sqrt(Ra*Pr)); no unit change applied\",\n  \"velocity\": \"identity for both (u_in = u_B in free-fall units, u_tgt = u_c); vel_factor = 1\",\n  \"temperature\": \"identity for both (stored as T/T0 in [0.4, 1.6], eps = 0.6); range asserted\",\n  \"pressure_deramp\": \"for each field fit p01(x,y) ~ a + b*y by least squares over the 64x64 field, y = row/(ny-1); r = p01 - (a + b*y) (zero-mean residual)\",\n  \"pressure_sign\": \"the generator's Boussinesq weak form uses '+ p*div(v)', the opposite of the paper's, so the raw Boussinesq pressure is the negative of the paper's P: P_in = -(p_raw_B - mean). The compressible member keeps the paper sign: P_tgt = +Ra*Pr*(p_raw_C - mean).\",\n  \"pressure_center\": \"spatial mean subtracted from P_in and P_tgt after scaling (m_in, m_tgt recorded)\",\n  \"approximation\": \"raw_both: none on the scale (both members exact); the pair is asymmetric because only the compressible member carries a hydrostatic ramp.  raw_both_deramp: the linear-in-y fit is applied identically to both, which is symmetric but also removes physical linear-in-y dynamic-pressure content from each member.\",\n  \"pressure_scale\": \"INPUT: the RAW Boussinesq solver pressure (P_bouss_raw.pt, units rho0*alpha^2/L^2) is used directly, s_B = 1, no ruler -- the corrected buoyancy Pr*Ra*(T-T0)/dT leaves no hydrostatic ramp.  TARGET: the RAW compressible solver pressure (P_comp_raw.pt, units rho0*U_c^2) is used directly too, s_c = 1 -- NO RULER ON EITHER SIDE.  The only operation on the target is the unit change k_c = Ra*Pr, which is exact rather than the per-sample (1/Fr^2)/|b_c| estimate the ramp ruler had to make from a min-max-normalised field. Both members are de-ramped with the same least-squares linear-in-y fit, which removes the compressible hydrostatic stratification and the (physical) linear-in-y part of both dynamic pressures.  This is the like-for-like variant, directly comparable with the v2 ramp_ruler and v3 raw_bouss_deramp datasets.\"\n },\n \"conversion_options\": {\n  \"min_abs_slope\": 0.1,\n  \"chunk\": 500\n },\n \"counts\": {\n  \"n_input\": 200,\n  \"n_kept\": 200,\n  \"n_dropped\": 0\n },\n \"dropped_samples_missing_Ra\": [],\n \"positive_slope_sample","encoding":"utf8","key":"test_ood/MANIFEST.json","name":"MANIFEST.json","size":320886,"truncated":true}],"sourceObjectKey":"test_in_distribution/per_sample.csv","starterCell":"import pandas as pd\ndf = pd.read_csv(\"hf://datasets/NurshatMenglik/NeuralConvection_2D_TrainTest/test_in_distribution/per_sample.csv\")\ndf.head()","starterSamplePath":"test_in_distribution/per_sample.csv","totalObjects":33},"summary":"Training and test tensors for the 2D Fourier Neural Operator (FNO) surrogate in \"A Neural Surrogate Approach for Simulating Natural Convection Problems\" (Menglik, Shao and Hyde,…","title":"NeuralConvection 2D: Boussinesq-to-Compressible Natural Convection Tensors for Fourier Neural Operators (Menglik, Shao, Hyde)","updatedAt":"2026-11-22T00:56:11.163Z","visibility":"public","websiteUrl":"https://huggingface.co/datasets/NurshatMenglik/NeuralConvection_2D_TrainTest"}}