Explanation · IllustrisTNG and SIMBA in CAMELS
How the two codes plant their black holes, and why it matters for the result
In plain words
The two codes put their first black holes into galaxies by different rules, and none of the four feedback dials changes that rule.
In both codes the black holes are put in by a rule: when a halo or a galaxy passes a size threshold, the code places one. IllustrisTNG plants a heavy one (8×105 h−1 M☉) in a halo above 5×1010 h−1 M☉. SIMBA plants a light one (104 h−1 M☉) in a galaxy above about 109.5 M☉ of stars. In the saved catalogues, at z = 0, 81% of IllustrisTNG galaxies in the lowest mass bin hold a black hole and 1% of SIMBA galaxies do. Fitted one property at a time, retuning the four dials can close 6 of the 11 properties studied, but not this one. This page explains one of our own results. No outside expert has reviewed the result or this explanation.
What this page is. An explanation of one internal result of this project (the all-at-once study). No outside expert has reviewed the result or this page. It does not claim that the result, or the explanation, is new: the CAMELS papers already report that the two codes differ, and the seeding rules below are the model authors' own. Every statement about how the codes seed black holes is a quotation from a paper or from a file of this repository, with its place; a statement that is in none of them is left out.
What seeding means
A galaxy simulation follows gas, stars and dark matter. The two models here also include central black holes. In both, the code places a black hole particle into the simulation by a rule, once a halo or a galaxy passes a size threshold. That step is called seeding. Here is how the authors of each model describe it.
In our cosmological simulations, a black hole with mass M_seed is placed at the centre of a halo whenever the on-the-fly friend-of-friends halo finder identifies a halo more massive than a threshold mass M_FOF that does not yet contain a black hole.
If a galaxy reaches a stellar mass M⋆ > γ_BH × M_seed and it does not already contain a black hole particle, then the star particle closest to the center of mass of the galaxy is converted into a black hole particle. For our fiducial simulations, we employ M_seed = 10⁴ M☉/h and γ_BH = 3×10⁵, which places black holes in galaxies with M⋆ ≳ 10^9.5 M☉.
The rule has two settings: how heavy the first black hole is (the seed mass), and how big a halo or galaxy must be before it gets one (the threshold). Both values are set by the model authors. The IllustrisTNG authors say of the seed mass:
Given the significant theoretical uncertainties in the early growth of SMBHs [citation omitted], we consider the seed mass as a poorly constrained free parameter.
The two rules side by side
The CAMELS paper describes both models as they were run in the experiment this study uses. For IllustrisTNG:
SMBH particles with initial mass M_seed = 8×10⁵ M☉ h⁻¹ are seeded in halos with mass M_FoF > 5×10¹⁰ M☉ h⁻¹.
And for SIMBA:
Black holes with initial mass M_seed = 10⁴ M☉ h⁻¹ are seeded in galaxies with M⋆ ≳ 10^9.5 M☉ by means of the on-the-fly FoF, with the M⋆ threshold for seeding motivated by FIRE simulations showing that stellar feedback strongly suppresses black hole growth in lower mass galaxies [citations omitted].
| IllustrisTNG | SIMBA | |
|---|---|---|
| What gets a black hole | a halo (a friends-of-friends group) | a galaxy, by its stellar mass |
| Threshold | halo mass above 5×1010 h−1 M☉ | stellar mass above about 109.5 M☉ |
| Seed mass | 8×105 h−1 M☉, which is 1.2×106 M☉ for h = 0.6711 | 104 h−1 M☉, which is 1.5×104 M☉ for h = 0.6711 |
The seed masses differ by a factor of 80. The value of h is the one the CAMELS paper fixes in all its simulations; the conversion is the only arithmetic on this page that is not a quotation.
… h=0.6711 …
Notice that one rule looks at the mass of the halo and the other at the stellar mass of the galaxy. The SIMBA authors give their reason for a threshold in stellar mass in the quotation above: it is motivated by other simulations. This page does not say why the IllustrisTNG authors chose theirs, because none of the sources quoted here does.
The four dials of the experiment are feedback parameters, not seeding parameters. The CAMELS paper says which parameters it varies:
In the hydrodynamic simulations, besides Ω_m, σ_8, and the initial random seed, we also vary the value of four astrophysical parameters related to stellar and AGN feedback. [...] While we do not consider parameter variations for other key baryonic processes, we explore variations over a broad range of efficiencies for some of the most important and uncertain feedback processes in galaxy formation: galactic winds driven by supernovae (SNe) and kinetic feedback from massive black holes.
The project's own protocol, written before the result existed, says the same thing in its own words:
Black hole occupation and mass compare seeding prescriptions (TNG seeds at 8e5/h Msun, SIMBA at much lighter masses), a discrete model choice that no dial changes, so the structural verdict of those families is expected by construction
What the saved catalogues show
If the rules are different, the catalogues should show it. They do. The figure counts, in each bin of stellar mass, the share of galaxies that hold a black hole.
In the lowest mass bin, 81% of IllustrisTNG galaxies hold a black hole and 1% of SIMBA galaxies do; in the next bin the shares are 83% and 4%. From log stellar mass 9.75 upward, at least 97% of SIMBA galaxies hold one, against 90% to 99% for IllustrisTNG. The pale bands show how far moving the dials, the cosmology and the random seed shifts each curve. In the lowest mass bin the 5th percentile over IllustrisTNG runs (54%) lies above the 95th percentile over SIMBA runs (4%), so no setting in these simulations closes that gap.
The CAMELS paper reports the matching feature in black hole masses:
We find good agreement between the median black hole masses of both simulation suites; however those for the SIMBA set drop quickly for stellar masses below ~5×10⁹ h⁻¹ M☉, owing to the black hole seeding model [citations omitted]; this feature is not observed in the IllustrisTNG set.
The protocol's input check of the two catalogues found the smallest non-zero black hole masses and the occupation in one simulation (LH_0, whose parameters are not the fiducial ones):
the smallest non-zero value is 1.2e6 Msun (TNG, seed 8e5/h) and 2.8e4 Msun (SIMBA), and the black hole occupation among M* > 1e9 is 0.88 and 0.33 at LH_0
The smallest IllustrisTNG value agrees, to the precision printed, with the paper's seed mass converted above. The smallest SIMBA value is larger than the paper's seed mass, and this page does not explain the difference.
What the study found
The study asked whether retuning each code's four dials can make the two agree on galaxy properties. Distances are in noise units: one unit is the scatter between one run of each code. Details are on the study page.
- All properties at once. The codes stay 4.8 noise units apart (half-sample range 4.6 to 4.9), far above the margin of 1.
- Black hole occupation is the largest single piece. Fitted alone, the two codes stay 9.4 apart at the best dial setting found, against 11.8 at the default dials (9.5 before the emulators' own error is removed). Its blocks carry 43% of the squared all-at-once distance. In the property's own unit, the codes' occupation curves differ by 0.64 (rms over the mass bins, fraction of galaxies) at the default dials and by 0.50 at the dial setting that brings them closest; one noise unit is 0.07.
- It is not the whole story. Without black hole occupation the all-at-once distance is 3.7; without occupation and black hole mass it is 3.8. Both stay above 1 (point estimates on all simulations).
- Black hole mass among galaxies that have one can be matched. Alone it is 0.8, below the margin. The dials can set the typical mass of a black hole; they cannot make the fraction of galaxies that hold one agree.
- Gas metallicity also stays apart, by 1.7 when fitted alone and with 25% of the all-at-once squared distance. It is not one of the two properties that compare seeding rules, and no quotation here explains it.
| Property | Fitted alone | Distance alone | Share of all-at-once | Compares seeding rules |
|---|---|---|---|---|
| Black hole occupation | structural | 9.4 | 43% | yes |
| Gas metallicity | structural | 1.7 | 25% | no |
| Stellar-to-halo mass ratio | unresolved | 1.5 | 2% | no |
| Stellar metallicity | unresolved | 1.1 | 4% | no |
| Star-forming main sequence | unresolved | 1.0 | 3% | no |
| Black hole mass, galaxies that have one | removable | 0.8 | 2% | yes |
| Stellar mass function | removable | 0.6 | 1% | no |
| Stellar half-mass radius | removable | 0.5 | 11% | no |
| Gas fraction | removable | 0.1 | 7% | no |
| Quenched fraction | removable | 0.0 | 0% | no |
| Stellar mass and SFR density | removable | 0.0 | 0% | no |
Distance alone: the best the dials of both codes can do for that property, in noise units, with the emulators' own error removed. Share: the part of the squared all-at-once distance that comes from the property's blocks at the best joint setting. "Compares seeding rules" is the protocol's own statement quoted in §2, not a result.
On the 97 outputs that all three codes provide (exploratory, after the registered run), the black hole occupation distance is 8.9 between IllustrisTNG and SIMBA, 2.0 between SIMBA and Astrid, and 0.0 between IllustrisTNG and Astrid. Taken alone, occupation separates SIMBA from the other two. Astrid's own seeding rule is not described in the sources this page quotes, so nothing is said about why.
What the study could not show
- That seeding causes the difference. No run changed a seeding rule. The experiment varies four feedback parameters and, in the words quoted in §2, does not vary other baryonic processes. The protocol records that a structural verdict for the two properties that compare seeding rules is "expected by construction", so the result for them says little beyond the fact that the rules differ. The analysis file adds that the codes also differ in other ways:
the codes differ in more than subgrid form: hydro solver, star formation algorithm, yields, wind treatment
- How much of the 4.8 would remain with the same rule in both codes. Leaving the occupation output out (3.7) is not the same as running both codes with one rule. Nobody has done that run here.
- Whether the difference is physical or comes from resolution. The experiment is at one resolution (25 Mpc/h boxes with 256³ particles of each kind). The repository's resolution note says:
black holes are planted by a rule that depends on resolution
- Dropping the two lowest stellar-mass bins of every property takes the all-at-once distance from 4.8 to 2.5. Those are the two bins where SIMBA's occupation is 1% and 4%, but the same cut applies to all properties, so the study does not say how much of the drop is occupation. No run at another resolution exists to test it.
- Anything about real black holes. The comparison is between two simulations, deliberately without observations. Neither rule is shown to be more realistic than the other.
- Anything about other codes. The third-code check above is exploratory, uses three codes from the one CAMELS project, and does not describe Astrid's rule.
- That the sources describe the runs exactly. The quotations are from the papers' descriptions of the models. They were not checked for this page against the configuration files of the CAMELS runs.
- Review. The registered protocol was reviewed internally only. No outside expert has reviewed the result, the quotations' use here, or this page.
Sources and how this page is made
The 12 quotations on this page are stored in data/seeding-sources.json, each with the paper or file and the place it comes from. For the papers, the entry also holds the sentence as it stands in the TeX source of the arXiv version. The papers were read on 2026-10-23:
- Villaescusa-Navarro et al. 2021, ApJ 915, 71 (the CAMELS paper), arXiv:2010.00619. It is the source named for the catalogues in data/ensembles/camels-galaxies.json, so it describes the two models as they were run in the experiment this study uses. SHA-256 of the source archive read:
d135837b37d56a586e54f09d4bb76cdb1d1cefaf032e59b8f7cd7808c2d6275e. - Weinberger et al. 2017 (the black hole model of IllustrisTNG), arXiv:1607.03486. It is listed as a paper of IllustrisTNG in data/simulation-profiles.json. SHA-256 of the source archive read:
7db366373c5be0ba002d0fbb1af0a1106655ba66571771ff952fdf0c08af84ca. - Davé et al. 2019, MNRAS 486, 2827 (the SIMBA paper), arXiv:1901.10203. It is listed as the overview paper of SIMBA in data/simulation-profiles.json. SHA-256 of the source archive read:
592fda461eb7915dda5add5069b6158bffa1ef5fcef17e9c9684f93be0634bd6.
The quotations from this repository are compared with the files word for word by the site test. The numbers on this page are computed by the site build from data/analysis/joint-reachability.json, data/retune/retune.json, data/analysis/third-code.json and data/ensembles/camels-galaxies.json; none is typed here. A reader who knows these codes can check each quotation against its paper, and the most useful correction would be a statement from the model authors about how the rules apply in the CAMELS runs.