Instrument · CAMELS · IllustrisTNG and SIMBA
Turn the physics dials
IllustrisTNG and SIMBA are two well-known simulated universes that disagree about how violently exploding stars and black holes push gas around. Each has four dials for it. Turn them, watch the galaxies change, and ask: can one universe be made to look like the other, and what stays apart?
Instrument
What am I looking at? (30 seconds)
- Each universe has four dials for how hard supernova winds and black holes push gas out of galaxies. Turn one and the 11 panels show how a galaxy property changes in both universes (orange: IllustrisTNG, blue: SIMBA).
- The sentence at the top counts how many of the 11 properties the two universes agree on. A shaded panel means they are further apart than two runs of the same universe with different random seeds.
- The buttons search the dials. Make SIMBA look like TNG matches the property chosen under Match on, then tells you what that did to everything else. You can also paste a curve of your own, or pick an observation, and see which dials reproduce it.
Loading the emulators…
How to read it
- Dials
- The four feedback strengths that the CAMELS simulations vary in both codes, each over a factor of 4 or 2 either way, shown as a multiplier on the default. The dials of the two codes are not the same physical thing: the first, for example, is the wind energy per unit of star formation in IllustrisTNG and the wind mass loading in SIMBA. The smaller grey line under each dial gives the technical name and what it means in that code. The cosmology is held fixed.
- Gap
- How far apart the two codes' curves are for one property, in units of the scatter you get between two runs of the same code that differ only in their random seed. Below 1 the codes agree as well as two seeds of one code would; a panel with a gap of 1 or more is shaded.
- Bands
- Narrow band: the seed-to-seed scatter at the current dials. Wide pale band: everything that moving all four dials can reach for that property alone (5th to 95th percentile).
- The buttons
- They search the dials for the smallest joint gap over the properties chosen in Match on, from the current setting and fourteen random starts, and then say in words what moved, what came into line, what was given up and what stayed apart. A setting that is good for one property and poor for others is the point: see the study.
- The panel that never closes
- Black hole occupation stays apart at every setting. Both codes put black holes in by a rule that none of the four dials changes; how the two rules differ, with the papers quoted.
- Grey lines
- Published curves from the dataset for the same property and epoch, one per source, without zoom-in samples or variants. Their definitions differ from the CAMELS catalogues (apertures, mass definitions, IMF, population cuts), so they are context, not a test. Hover for the source and definitions; dotted lines are digitized from figures.
Basis and limits
Each curve is a polynomial emulator (degree 2 or 3, ridge-regularised, weighted towards the fiducial cosmology) fitted to 1,000 Latin-hypercube simulations per code and checked against 27 cosmic-variance seeds; the coefficients are in retune.json and nothing is fitted in the page. The observables are binned summaries of the Subfind catalogues of 2,054 CAMELS simulations at z = 0, 0.95, 2.15, 3 and 4 (data/ensembles/camels-galaxies.json). Both codes are read from catalogues made by the same Subfind code with the same field definitions, so no aperture or IMF difference separates them. The codes still differ in solver, star formation algorithm and yields, not only in their feedback dials.
How well the emulators predict simulations they have not seen
Each simulation within 0.5 of the fiducial cosmology (245 (IllustrisTNG) and 255 (SIMBA) of them) was predicted by an emulator fitted without the tenth of the simulations it belongs to. The table gives the median error over the outputs of each property, in the property's own units, and as a multiple of the scatter between two runs of the same code (which no emulator can beat). A ratio near 1 means the emulator is as good as the simulations allow; the larger ones, such as stellar half-mass radius and the stellar mass function, are around 1.3 to 1.4.
| Property | unit | IllustrisTNG | SIMBA | ||
|---|---|---|---|---|---|
| median error | × seed scatter | median error | × seed scatter | ||
| stellar mass function | dex | 0.091 | 1.35 | 0.089 | 1.31 |
| quenched fraction | fraction | 0.054 | 0.95 | 0.046 | 0.97 |
| star-forming main sequence | dex | 0.067 | 1.36 | 0.059 | 1.19 |
| black hole mass at fixed M* | dex | 0.100 | 1.19 | 0.079 | 1.44 |
| black hole occupation | fraction | 0.051 | 1.11 | 0.024 | 1.08 |
| gas metallicity at fixed M* | dex | 0.030 | 1.35 | 0.028 | 1.20 |
| stellar metallicity at fixed M* | dex | 0.019 | 0.93 | 0.032 | 1.38 |
| stellar half-mass radius at fixed M* | dex | 0.030 | 1.34 | 0.031 | 1.33 |
| gas fraction at fixed M* | fraction | 0.026 | 1.09 | 0.026 | 1.11 |
| stellar-to-halo mass ratio | dex | 0.047 | 1.17 | 0.044 | 1.15 |
| stellar mass and SFR density | dex | 0.072 | 1.07 | 0.078 | 1.05 |
- CAMELS boxes are 25 Mpc/h with 256³ dark matter particles plus 256³ gas elements: few galaxies above 1011 M☉, a lower mass limit of 109 M☉ (about 60 star particles), and almost nothing usable at z ≥ 3. Everything here is for that resolution.
- The dials reach the corners of the design, where an emulator trained on a Latin hypercube is least constrained; the study's controls share that search.
- Two codes only. What separates these two says nothing yet about other models.
The study behind it, with its protocol written before the result and a calibration on synthetic worlds: PROTOCOL-joint-reachability.md. The page state is in the address (dials and redshift), so a setting can be shared as a link.