The Dataset Viewer has been disabled on this dataset.

Capsule forebody aeroheating, Newtonian and Navier–Stokes CFD

The capsule's CAD, TopoLink mesh and heat flux morphing together from a Varda-like to a Gemini-like, an Orion-like and a Mars-entry-like shape

The capsule family with Newtonian heating, meshed like the dataset's walls; an illustration at Mach 10 and 60.6 km.

Laminar, perfect-gas heat flux and pressure on the heat shield of a parametric entry capsule, at two fidelities. There are 1,024 samples, each a unique shape and flight condition, from a Varda-like 0.8 m to beyond Orion's 5 m and from Mach 6 to 15. Every capsule has a solution from the Newtonian panel method, a fast engineering model, and 64 of them also have a second-order laminar Navier–Stokes CFD solution from Eilmer. The meshes were created using TopoLink, a topology-linked, hybrid-cell, structured mesh generator. Every wall has the same nodes, so node i is the same point of the heat shield on every capsule and at both fidelities. Every run also carries what is needed to estimate its heating on a hotter wall, and each CFD case comes with its whole flow field around the heat shield.

Air temperature around one capsule from SU2's first-order start and Eilmer's converged solution, and the wall heat flux of the three methods

One CFD case (capsule A below) at Mach 13.2 and 53 km: SU2 (Eilmer's starting field), Eilmer on its shock-fitted grid, and the wall heat flux of all three methods.

The dataset is stored in Gamut files, readable from Python, MATLAB, Julia, C++, C, Rust and the command line: vehicle_capsule_multifidelity.gamut on the triangulated wall, vehicle_capsule_multifidelity_quad.gamut with the same runs on the wall's own quads, and vehicle_capsule_eilmer_volume.gamut with the CFD cases' flow fields.

At a glance

Members 1,024 sampled capsules (label member), each one shape at one flight condition
Runs 1,088: 1,024 Newtonian (fidelity = newtonian_panel) and 64 Eilmer (fidelity = eilmer_second_order)
Inputs diameter, nose radius, forebody angle and shoulder radius (shape); Mach number and altitude (flight)
Wall each capsule's own TopoLink CFD wall, the full heat shield to mid-shoulder: 2,361 nodes and 4,640 triangles, or 2,320 quads in the quad file, the same on every run; coordinates in metres at real size
Fields heat flux and pressure at the nodes; heat flux on the wall faces; arc length from the stagnation point; for a hotter wall, the boundary-layer edge state, recovery enthalpy and film coefficient at the nodes
Flow fields for each CFD case, density, velocity, pressure, temperature and Mach number in 232,000 cells from the wall to just outside the bow shock, numbered after the wall
Scalars per run heat load (the heat flux integrated over the wall at the run's flight condition, in W; not over a trajectory); stagnation, peak, shoulder and mean heat flux; stagnation pressure; forebody pressure-drag coefficient (pressure only, heat shield to mid-shoulder); freestream state and total enthalpy
Ranges, Newtonian stagnation heat flux 2.49–1,884 kW/m², heat load 2.89–8,705 kW
Ranges, Eilmer stagnation heat flux 8.47–738 kW/m², heat load 15.1–8,795 kW
Held fixed calorically perfect air, laminar flow, a 300 K wall, zero angle of attack
Files vehicle_capsule_multifidelity.gamut (triangles), 80.4 MB; vehicle_capsule_multifidelity_quad.gamut (quads), 75.4 MB; and vehicle_capsule_eilmer_volume.gamut (flow fields), 470 MB. gamut info lists their contents, and their source attribute describes the models and checks

Two fidelities on one wall

The Newtonian panel method (fidelity = newtonian_panel) is an engineering model: modified Newtonian pressure, Fay–Riddell stagnation-point heating (its perfect-gas form) at the velocity gradient of the nose radius, and the Lees laminar distribution elsewhere. A run takes seconds, so every capsule has one. It cannot see how the shoulder changes the flow over the face, and it under-predicts the heating peak on the shoulder.

Eilmer (fidelity = eilmer_second_order) is CFD: the laminar Navier–Stokes equations, second order, on a 232,000-cell grid fitted to the bow shock around the whole heat shield. A case took a median of 1.9 hours of wall time on six cores (1.2–5.4 hours), including the SU2 start and any time waiting to start, so only a sample of the capsules has one.

The cheap model maps the heating over the whole design; the expensive one shows how far off it is, and where. A field such as heat_flux is one array of runs × nodes, so the difference between the fidelities is a subtraction at each node, with no interpolation between meshes.

How good is each fidelity

Along the heat shield of three of the CFD cases:

Heat flux along the heat shield of three different capsules, the Newtonian panel method against Eilmer on the same nodes

Both fidelities on the same nodes; the shoulder is grey, and Eilmer's curve is a running median over 25 nodes.

On all 64 CFD cases, the Newtonian value minus Eilmer's, as a percentage of Eilmer's:

The Newtonian method minus Eilmer, in per cent, for heat load, stagnation heat flux and shoulder peak heat flux, against nose radius

One point per CFD case; A, B and C are the capsules above.

  • Heat load. Below Eilmer's on every case, by 1.0–15.4 % (median 5.9 %), most on the largest noses: the 6 of 1.2 D and above are 11.2–15.4 % low, and the four furthest off are among them. The closest 9 are less than 2 % low; the uncertainty in Eilmer's own heat load was not measured.
  • Stagnation heat flux. From +40.3 % (capsule_0679, a 0.29 D nose on a 75° cone) to −13.0 % (capsule_0066, a 1.27 D nose). It is low on 14 of the 15 noses above 0.9 D, by up to 13.0 %, and 2.9 % high on the other (capsule_0062, a 0.92 D nose); on smaller noses it is high or low depending on the face. This is consistent with the panel method taking the stagnation velocity gradient from the nose radius alone, where in the flow the whole forebody sets it.
  • Shoulder peak. The panel method's highest heat flux on the shoulder (from the ring of nodes where the cone meets it) is 11–43 % below Eilmer's (median 24 %). By node values, the shoulder is the hottest point of the heat shield on 21 cases in Eilmer and on 6 in the panel method.
  • Pressure drag. The panel method's forebody pressure-drag coefficient is −8.9 to +10.4 % off Eilmer's (median +4.3 %). It reads high on 48 cases, more so the larger the forebody angle, and low only on forebody angles up to 65.2°.
  • Eilmer itself. At zero incidence the heating should be the same all the way round the axis. Over 3–80 % of the arc, Eilmer's wall faces depart from the median at their distance from the stagnation point by a standard deviation of at most 1.20 % on any case.

Heating on a hotter wall

Every run has a 300 K wall. For a hotter one, each run carries at every node:

Field Meaning
edge_temperature, edge_enthalpy, edge_velocity, edge_density, edge_mach_number the gas at the edge of the boundary layer, estimated from the node's pressure: the gas behind a normal shock, expanded isentropically to that pressure
recovery_enthalpy h_r = h_e + ½ √Pr u_e², the enthalpy an insulated wall would reach (laminar)
film_coefficient C_H = q / (h_r − h_w), from the node's heat flux q and the 300 K wall's enthalpy h_w (the constant wall_enthalpy)

Checked against the boundary-layer edge of all 64 Eilmer solutions, the recovery enthalpy agrees within about 1 % over the whole wall (5–95 % of wall faces). The edge state itself is approximate. The edge velocity is within a few per cent on the nose and cone where the boundary layer is thin. On the cone, where a thick boundary layer takes in the entropy layer, it reads low, by up to 18 % (5th percentile of faces; up to 26 % on single cases). On the shoulder, where the wall pressure falls below the pressure at the boundary-layer edge, it reads 3–27 % high (5–95 % of faces).

The heat flux on a wall at temperature T_w is then approximately

q(T_w) ≈ C_H (h_r − c_p T_w), c_p = 1004.675 J/(kg K),

and a radiative-equilibrium wall temperature solves C_H (h_r − c_p T_w) = ε σ T_w⁴ for an emissivity ε. At the stagnation point a 1,500 K wall keeps 11–90 % of the 300 K heating, least at the lowest total enthalpy (Mach 6.0) and most at the highest (Mach 15.0).

The edge state is estimated the way the CBAero engineering code estimates it. Holding C_H at its 300 K value slightly overpredicts the heating of a hot wall, because the film coefficient falls a little as the wall heats. With three CFD cases rerun in Eilmer at 1,000 and 1,500 K walls, this correction gave 1.8–2.9 % (1,000 K) and 2.9–4.8 % (1,500 K) more heat load ahead of the shoulder than Eilmer, and 1.8–4.4 % more stagnation heat flux, close to what the exact stagnation-point solution and the Fay–Riddell wall term predict. That check does not cover the shoulder, dissociating or catalytic air, or turbulent flow.

Flow fields

vehicle_capsule_eilmer_volume.gamut holds Eilmer's converged flow field for each of the 64 CFD cases. These are the solutions of the Eilmer runs in the wall files, at the same solver step, and each run carries the same inputs and freestream values; join the files on member.

  • Grid. The wall's 2,320 quads are extruded through 100 layers of cells to a boundary just outside the bow shock, where the freestream is imposed: 232,000 hexahedra and 238,461 nodes. Every case has the same cells, with its own node positions.
  • Numbering. Nodes and cells follow the wall files. The first 2,361 nodes are the wall files' nodes, in the same order and at the same positions, and node j + 2,361 l is node j carried l layers out. Cell f + 2,320 l is in layer l over wall face f, cell f of the quad file. A cell field reshaped to (100, 2,320) therefore holds each wall face's column of cells, from the wall to the inflow boundary.
  • Fields. Eilmer's own cell values, in single precision: density, velocity_x, velocity_y, velocity_z, pressure, temperature and mach_number. x runs along the axis in the flow direction, with the nose at the origin.
  • Boundaries. The node label boundary marks the wall, the inflow boundary, the outflow boundary and the interior. The wall ends at mid-shoulder, where the outflow boundary starts, so there is no wake or afterbody flow.

Inputs

Name Range Units Meaning
diameter 0.8–5.5 m maximum diameter D, sampled log-uniform
nose_radius 0.25–1.6 × D radius of the spherical nose cap
forebody_angle 55–75 degree half-angle of the cone between the nose cap and the shoulder
shoulder_radius 0.025–0.075 × D radius of the toroidal shoulder
mach 6–15 1 freestream Mach number
altitude 40–80 km freestream from the 1976 US Standard Atmosphere

Held fixed for every run: zero angle of attack, a 300 K isothermal wall, and calorically perfect air (γ = 1.4, R = 287.05 J/(kg K), Sutherland viscosity, Pr = 0.72), laminar.

The 1,024 capsules in three projections of the six inputs, with the CFD cases marked

The capsules are a Latin hypercube over the six inputs, with the nose radius drawn as a fraction of the largest buildable at each forebody angle and shoulder radius (nose_radius_fraction); that limit is the empty corner of the middle panel. The CFD cases come from a 64-capsule subset of the design. The first 32 are matched, by optimal assignment, to the points of a 32-point Latin hypercube over the same inputs. The next 29 come from the rest of the subset in a farthest-first order, in which each capsule is the farthest in the six inputs from all the capsules before it. Three in the subset could not be run by this pipeline: two shapes whose CFD grids could not be extruded, and one at Mach 14.6 and 44 km whose starting solution diverged. Their places went to the next capsules of the whole design in the same order, after three more whose starting solutions did not converge (one at Mach 14.9 and 64 km, and two of the largest capsules at 80 km). Together the CFD cases spread over the whole design but are not themselves a Latin hypercube.

Quick start

pip install bijective-gamut
hf download bijective/capsule-aeroheating --repo-type dataset --local-dir capsule
gamut info capsule/vehicle_capsule_multifidelity.gamut

In Python, choosing runs by fidelity and pairing each Eilmer run with the Newtonian run of the same member:

import numpy as np
import gamut

with gamut.open("capsule/vehicle_capsule_multifidelity.gamut") as study:
    runs = study.table()                                    # inputs, results and labels, one entry per run
    cfd = study.select("fidelity == eilmer_second_order")   # positions of the Eilmer runs
    newt = study.select("fidelity == newtonian_panel")      # positions of the Newtonian runs, one per member
    row_of = dict(zip(runs["member"][newt], newt))
    pair = np.array([row_of[m] for m in runs["member"][cfd]])   # each Eilmer run's Newtonian run
    q_cfd = study.get_field("heat_flux", rows=cfd)          # (Eilmer runs, nodes), W m-2
    q_newt = study.get_field("heat_flux", rows=pair)        # the same members, node for node
    C_H = study.get_field("film_coefficient", rows=cfd)     # kg m-2 s-1
    h_r = study.get_field("recovery_enthalpy", rows=cfd)    # J kg-1
    xyz = study.get_points(int(cfd[0]))                     # wall nodes of the first CFD member, m
    cells = study.get_cells()                               # the shared triangles, as VTK gives them

difference = q_newt / q_cfd - 1        # Newtonian against Eilmer at every node of every CFD case
q_1500 = C_H * (h_r - 1004.675 * 1500)  # Eilmer's heating on a 1,500 K wall, approximately (a few % high)

The flow fields, one run per CFD case:

with gamut.open("capsule/vehicle_capsule_eilmer_volume.gamut") as flow:
    member = flow.table()["member"]                         # the CFD case of each run
    T = flow.get_field("temperature", rows=0)               # (cells,), K, the first case
    columns = T.reshape(100, -1)                            # [layer, wall face]; layer 0 touches the wall
    xyz = flow.get_points(0)                                # its nodes, m; the first 2,361 are the wall's
    hexes = flow.get_cells()                                # the shared hexahedra, as VTK gives them

From the command line, without writing code:

# the Eilmer runs and the Newtonian runs as two CSV tables; join them on member
gamut get capsule/vehicle_capsule_multifidelity.gamut member diameter nose_radius mach altitude heat_load stagnation_heat_flux --where "fidelity == eilmer_second_order" -o eilmer.csv
gamut get capsule/vehicle_capsule_multifidelity.gamut member diameter nose_radius mach altitude heat_load stagnation_heat_flux --where "fidelity == newtonian_panel" -o newtonian.csv

# a smaller study file of the Eilmer runs alone, then one .vtu per run, with every field, for ParaView
gamut get capsule/vehicle_capsule_multifidelity.gamut --where "fidelity == eilmer_second_order" -o eilmer.gamut
pip install "bijective-gamut[vtk]"
gamut-vtk export eilmer.gamut eilmer_vtk/

gamut-vtk export capsule/vehicle_capsule_multifidelity.gamut capsule_vtk/ writes every run the same way, and gamut-vtk export capsule/vehicle_capsule_eilmer_volume.gamut flow_vtk/ writes each CFD case's flow field (about 18 MB a case). To download without the flow fields, add --exclude "*volume*" to hf download. Gamut has libraries for MATLAB, Julia, C++, C and Rust as well; see the documentation.

How it was made

Each capsule's CAD comes from the vehicle_capsule parametric family and its wall mesh from TopoLink, with the same nodes and triangles on every capsule. The Newtonian panel method ran on the wall of every capsule. For each CFD case, a first-order laminar SU2 solution on a grid extruded from the wall served only as Eilmer's starting field: it reads the stagnation heat flux 4.8–21.0 % above Eilmer's and is not in the file. Eilmer then ran on two grids fitted to the shock: on the first until its CFL number reached 1,000 or for about 150 steps, on the second until the wall heat flux and the heat load were steady. On 7 of the 8 cases whose shock sat more than 1 % inside the second grid's fitted layer on the axis, it then ran on a third grid to the same test (on the eighth, the third grid fitted the shock less well and was not used). The flow-field figure near the top shows one case, and review/cfd_case_review.pdf has a page for every CFD case: its converged solution (wall heat flux, and the Mach number on the symmetry plane), its convergence from the SU2 start through its Eilmer grids, and its checks, after a table of all the cases.

Limitations

  • The Newtonian method's errors. Its heat load is low on every CFD case, its stagnation heat flux is high or low depending on the shape, and it under-predicts the shoulder peak.
  • Few CFD cases. 64 cases over six inputs show the size and pattern of the Newtonian errors, not the CFD heating everywhere in the design.
  • Skipped cases. Six selected CFD cases could not be run (see Inputs), so the CFD cases are sparser near those shapes, near Mach 15 and near 80 km for the largest capsules.
  • Hot-wall heating is approximate. A film coefficient held at its 300 K value runs a few per cent high on a hot wall (see Heating on a hotter wall).
  • No error bars on the CFD. Eilmer's results rest on its grids and convergence checks; the grid was not refined case by case.
  • Simplified physics. Calorically perfect gas (no chemistry or vibrational excitation, which matter over most of the Mach range, increasingly toward its top), laminar flow, zero incidence, a 300 K wall, and the heat shield alone, with no afterbody or wake.

Versions

  • Version 4 (16 November 2026) adds 11 CFD cases, 64 in all; hot-wall quantities on every run (the boundary-layer edge state, recovery enthalpy, film coefficient and total enthalpy); vehicle_capsule_multifidelity_quad.gamut, the same runs on the wall's own quads; and vehicle_capsule_eilmer_volume.gamut, the flow field of every CFD case. 7 CFD cases now use a third grid fitted to the shock, 5 of them from version 3: their stagnation heat flux moved by −0.9 to +4.9 %, by more than 2 % on 3 (capsule_0066 (+4.9 %), capsule_0745 (+2.7 %), capsule_0565 (+2.3 %)), and their heat load by at most 0.062 %. Version 3's other runs are unchanged.
  • Version 3 (14 November 2026) added 21 CFD cases, 53 in all, and two results on every run: shoulder_heat_flux and forebody_pressure_drag_coefficient.
  • Version 2 (11 November 2026) added vehicle_capsule_multifidelity.gamut with 32 CFD cases.
  • Version 1 (4 November 2026), vehicle_capsule_newtonian.gamut, is unchanged: the same 1,024 capsules with the Newtonian panel method alone, on TopoLink's family mesh of 2,929 nodes and 2,880 quads. The two files' Newtonian heat loads agree within 0.3 % and their stagnation heat fluxes are the same. Version 2 adds the label fidelity, the field heat_flux_cell and the result stagnation_pressure, and drops version 1's label kind and result stagnation_velocity_gradient. Version 1's card and figures are under the tag v1.

Each version is tagged in this repository (v1 to v4). For an earlier one, add --revision v3, for example, to hf download.

Licence and citation

The data and figures are licensed under CC BY 4.0.

@misc{bijective_capsule_aeroheating_2026_v4,
  title        = {Capsule forebody aeroheating, Newtonian and Navier--Stokes {CFD}},
  author       = {{Bijective Systems}},
  year         = {2026},
  version      = {4},
  howpublished = {\url{https://huggingface.co/datasets/bijective/capsule-aeroheating}}
}

Version 1 alone (vehicle_capsule_newtonian.gamut) keeps its original entry, bijective_capsule_aeroheating_2026, titled "Capsule forebody aeroheating, engineering model (demo)".

About

Made by Bijective Systems with TopoLink (meshes for parametric geometry families) and Gamut (an open file format for parametric simulation studies). The CFD used Eilmer, with starting fields from SU2. Proximant, a Bijective Systems product, chose the design and fits surrogate models to studies like this one.

Downloads last month
522