Star formation in the Galaxy and Across Cosmic Time

PLACEHOLDER — NOT PART OF THE TALK

With JWST, ALMA and the VLA we can now resolve individual forming stars, their outflows and even their ice chemistry across the Milky Way’s Central Molecular Zone — and the clouds there turn out not to be alike.

That census still covers only a small fraction of the CMZ; the JWST Treasury survey and Roman will finish it and, for the first time, place these clouds in three dimensions.

What the audience will actually hear, read off the slide order — see the comment above for where this differs from the intended narrative.

Slide 0 - Instructions:

To move slides, use the arrow keys or swipe on your mobile device
To see the speaker notes, press "s"
To go to full screen, press "f"
To print as PDF, go to this URL: ?print-pdf, then print.
To get a PDF with speaker notes, add ?print-pdf&showNotes=true to the URL.

What sets star and planet formation in the Galaxy’s richest environments across cosmic time

Postdocs: Nazar Budaiev (2026-), Theo Richardson (2025-2026), Miriam Garcia Santa Maria (2024-2025), Allison Towner (2020-2023)
PhD: Desmond Jeff (2025), Theo Richardson (2025), Alyssa Bulatek (2026), Nazar Budaiev (2026), Savannah Gramze (2027), Taehwa Yoo (2028)
Undergrad/postbac: Derod Deal, Aden Dawson, Ethan Bhula, Mario Antonio Daley, Prashant Sikhdar, Avery Lacon, Laya Damaraju
Supported by NSF 2008101, 2206511, CAREER 2142300, STSCI 1905, 2221, 3523, 5365, 6151, 10662, 10678, Roman 19008, 19012, Astropy

Slides available at https://keflavich.github.io/talks/colloquium_sep2026_granada.html or from my webpage →talks

The Galactic Center is where we study cosmologically relevant conditions at planet-forming scales

Part 1: We see star formation in the CMZ, but it's different.
Part 2: CMZ clouds are icier and dustier
Part 3: Future: JWST-GC, Roman, tomography

What happens to
proto-stellar discs
in these extreme
environments?

Proplyds: discs photoevaporated

Accretion-driven truncation
Wijnen+ 2017; Otter+ 2021

Madau & Dickinson 2014

Most of the stars in the universe formed at cosmic noon, z ~ 1-3
Those conditions -- density, UV field, turbulence -- are what the CMZ reproduces locally
So: what happens to the discs, and to the planets, under those conditions?
Credit the truncation argument correctly: the mechanism is Wijnen+ 2016, 2017 (Pols, Pelupessy, Portegies Zwart — Nijmegen/Leiden): face-on accretion of ISM material carrying no azimuthal angular momentum shrinks the disc; ram-pressure stripping removes the rest. It shows up in the observations via Otter+ 2021 (Otter, Ginsburg, Ballering, Bally+) — the OMC1 discs are embedded, hidden from the Trapezium's UV, and still small, so photoevaporation cannot be what truncated them. Moeckel & Throop 2009 (Bondi–Hoyle–Lyttleton accretion onto a protoplanetary disc) made essentially the same argument a decade earlier, as did Bally. The picture on the slide is Bally+ 2020's; the physics is not his.

The CMZ

$\sim10^8$ M$_\odot$ of gas in $\sim200$ pc, 10% of Galactic star formation

Image: NRAO, Ginsburg

Ginsburg+ 2018a

$>\frac{1}{3}$ of CMZ SF is in bound clusters (3-8$\times$ local)

$\sim$50% of CMZ SF occurs in the Sgr B2 cloud.

Barnes+ 2017, Hatchfield+ 2024, Ginsburg+ 2018b, Budaiev+ 2024

Zoom-in to the Sgr B2 cloud:
Most massive, actively star-forming cloud in the galaxy: 10⁷ M_⊙
ALMA 3mm image showing mix of dust and free-free emission
Where the ~50% comes from, if anyone asks:
- Barnes+ 2017 (MNRAS, "Star formation rates and efficiencies in the Galactic Centre"): CMZ total SFR = 0.09 ± 0.02 M_⊙/yr averaged over IR-luminosity, YSO-counting and free-free methods; their Sgr B2 entry is 0.005 M_⊙/yr from the IR, rising to 0.045 when the embedded stellar mass is taken from VLA+WMAP instead — i.e. half the CMZ total.
- Hatchfield+ 2024 (CMZoom IV): "we find a total SFR (excluding the CND) of ~0.05 M_⊙/yr, the majority of which is due to the estimate of the SFR of Sgr B2 (~0.036 M_⊙/yr)" — that is ~70%, so 50% is the conservative statement.
- Ginsburg+ 2018b ("Distributed star formation throughout the Galactic Center cloud Sgr B2") is where the 0.036 M_⊙/yr itself comes from — it is the number Hatchfield adopts.
- So the two independent totals (0.09 and 0.05) bracket the answer and both put Sgr B2 at half or more. The honest caveat is that Sgr B2's SFR depends on how you count embedded stellar mass; the IR-only route gives ~10x less.

The Central Molecular Zone of the Galaxy represents one extreme of star forming conditions in the Galaxy

PPVII review: Henshaw, Barnes, Battersby, Ginsburg, Sormani, Walker

Sgr B2 with JWST/ NIRCam MIRI on MEERKAT: Ionized gas shows HII regions (OB stars)

Budaiev+ 2026

Protoclusters

MIRI reveals deeply embedded feedback sources

Budaiev+ 2026

MIRI 25 micron shows the outflow:
first IR light from within Sgr B2 N;

$\dot{M} \sim 10^{-1} \mathrm{~M}_\odot \mathrm{~yr}^{-1}$, $M_{cl} \sim 3\times10^3 \mathrm{M}_\odot$

Budaiev+ 2026

Sgr B2 N Accretion + Outflow

Budaiev+ 2025: H$_2$O masers & SiO outflow

Schwörer+ 2019: Accretion along filaments: $\sim0.1~\mathrm{M}_\odot \mathrm{yr}^{-1}$

Massive clusters account for 30-40% of the star formation

A dark, dusty swath hides ongoing unclustered SF

Sgr B2 Deep South — extended star formation with JWST

Sgr B2 Deep South — mostly seen by ALMA

NIRCam starless
F480M (hot dust &/or CO+H₂) Brα Paα

ALMA 1.3 mm continuum
contour: ALMA SiO 5–4

Sgr B2 Deep South — Bipolar flow; 5$\mu$m from the cavity?

NIRCam starless
F480M (hot dust &/or CO+H₂)Brα Paα

ALMA 1.3 mm continuum
contour: ALMA SiO 5–4

Sgr B2 Deep South — DS9, a hot core containing a massive star, along a thin filament

NIRCam starless
F480M (hot dust &/or CO+H₂) Brα Paα

ALMA 1.3 mm continuum
contour: ALMA SiO 5–4

Zooming back out to the overview...

JWST finds the extended, older star formation

ALMA finds the embedded, ongoing star formation

◯ JWST 4.8μm excess sources (Budaiev+ 2026)

■ ALMA 3 mm continuum sources (Ginsburg+ 2018, Budaiev+ 2024, Daley+ in prep)

━ JWST 4.8μm excess source density

There is an overall asymmetry in the star formation seen both in JWST (HII regions) and ALMA (embedded YSOs)

JWST shows the transition is sharp

Budaiev+ 2026

What ALMA sees, JWST doesn't: 3/700 point source matches
(HII regions, outflow cavities do match)

Star formation is ongoing
JWST & ALMA see different phases - complementary
Sgr B2 asymmetry is compression/collision

Zooming back out to the Galactic Center as seen by ACES....

JWST has observed much of the Galactic Center

The Galactic Center in the mid-infrared (Spitzer/GLIMPSE)

…and in the radio

…with ACES + MUSTANG on top

Sgr B2: >500 YSOs w/ALMA
~100 HII regions JWST+ALMA+VLA

Ginsburg+ 2018, Jeff+ 2024, Budaiev+ 2024, 2025, 2026, Daley+ in prep

Cloud E/F: $10^5\mathrm{~M}_\odot$ gas, 1? YSO

Houghton, Esteve, Gutermuth, Longmore, Huard, Kauffmann subm.

Clouds C & D: $10^5\mathrm{~M}_\odot$ gas, one small “cluster” of 5–10 YSOs each

Gramze+ in prep

The Brick: $10^5\mathrm{~M}_\odot$ gas, about 10 YSOs, one above 5 M$_\odot$

Walker+ 2021

Quintuplet & Arches: no gas, few-Myr old, >10$^4$ M$_\odot$ clusters

Hosek+ 2019

Sgr C: $10^4\mathrm{~M}_\odot$ gas, ~5 IR–ALMA matched YSOs,
~40 inferred from outflows

Crowe+ 2023, Lu+ 2021

Star formation in the CMZ does not follow the local relation

Nearby clouds all follow one relation, Sigma_SFR ~ Sigma_gas^2.0, with a uniform efficiency per free-fall time (eps_ff ~ 0.026)
Say the caveat out loud: Pokhrel+ fitted this only below 676 Msun/pc^2 (their N(H2) < 3e22 cut), and their own Fig. 3 shows eps_ff declining above ~1000. Every CMZ point here is 1–2 dex beyond the calibration — pure extrapolation, exactly as I said about Gutermuth in the 2018 paper
The bands are cloud-to-cloud scatter: 0.30 dex, so 2 sigma = 0.60 dex

Where each CMZ point comes from — all three limits drawn as generously as the data allow, so the gap is not overstated:

Sgr B2 (Ginsburg+ 2018) — the whole Fig. 17a point cloud, not one point. Fig. 17a plots Sigma_* vs Sigma_gas, so the whole panel is divided by an age to get Sigma_SFR. 556 points digitised straight out of the paper's raster. Age = 0.5 Myr is used: it matches Pokhrel's t_PS so both axes mean the same thing, and it is more conservative than the paper's own 0.74 Myr (a shorter age raises the rate). Median (log Sigma_gas, log Sigma_SFR) = (3.81, 2.08); the relation predicts 3.52, so 1.44 dex below. Age sensitivity, all still below: 1 Myr → 1.74 dex below; 0.5 Myr → 1.44; 0.1 Myr (the extreme) → 0.74 dex below, still outside 2 sigma. So the conclusion does not depend on the age. Axis calibration was checked against the figure's own Gutermuth alpha=1 model lines: at Sigma_gas = 1e4 the t = 0.74 and 0.01 Myr curves are predicted at page rows 678 and 989 and found at 677 and 993. Caveats to own: Fig. 17a assumes 95 Msun per source (21.8 Msun mean x 23% IMF sampling), the sample misses Class II sources so the stellar mass may be underestimated by up to 2x, and the right-hand column of saturated-Herschel lower limits is excluded from the cloud drawn here.
The Brick (Walker+ 2021), log Sigma_gas = 3.71, limit log Sigma_SFR < 1.03 — 2.28 dex below the relation, and ~1.3 dex below Sgr B2 at the same gas surface density. Cloud-scale aperture, to match how Pokhrel measure and so the two limits are comparable: >1e5 Msun within R = 2–3 pc (Sec 1); R = 2.5 pc → 19.6 pc^2 → 5.1e3 Msun/pc^2. Walker+ quote no SFR and no limit, so the ceiling is ours and deliberately generous: every Table 3 core at its 22 K mass, source 1 counted whole (64.2 Msun) not split, = 104 Msun of core gas, times a core-to-star efficiency of 1.0 — every gram already a star — / 0.5 Myr / 19.6 pc^2 → 10.6. These cores really are protostellar: 9 of 18 drive SiO outflows. Using the 19" ALMA field alone instead (7e3 Msun, 1.75 pc^2) moves it to (4.0e3, 119) — same cloud, aperture ten times smaller.
Cloud E/F (Barnes+ 2019 + Walker+ 2018), log Sigma_gas = 3.78, limit log Sigma_SFR < 0.29 — 3.16 dex below the relation, ~2 dex below Sgr B2. Sigma_gas from Walker+ 2018 Table 1, which tabulates the dust ridge on a consistent footing: cloud e = 11.2e4 Msun in R = 2.4 pc, cloud f = 7.3e4 Msun in R = 2.0 pc; Barnes treat them as one cloud, so 1.85e5 Msun over 30.7 pc^2 = 6.0e3 Msun/pc^2. The Immer+ 2012 column of the same table gives 2.6e3, so the horizontal position is good to a factor of ~2. The ceiling is set by what is observed: no HII regions, no protostellar source catalogue, and the only star formation signature in either cloud is one H2O + Class II CH3OH maser pair in the south. A Class II methanol maser marks a high-mass protostar, so allow one plus companions, ≤ 30 Msun — generous, and bounded above by the Brick's entire counted protostellar content (104 Msun in 18 cores), since E/F shows strictly less. / 0.5 Myr / 30.7 pc^2 → 2.0.
The horizontal bars are the range of column density each cloud actually covers: from the CMZoom threshold N(H2) = 1e23 cm^-2 (Walker+ 2018 Sec 2.1, "designed to target all regions within the CMZ that lie above a column density threshold of ~1e23 cm^-2") = 2.2e3 Msun/pc^2, up to the densest structure in each cloud — the Brick's maser core, 72 Msun in R = 0.04 pc (Rathborne+ 2015 via Walker+ 2021) = 1.4e4 = 6.4e23 cm^-2; and Cloud E/F's peak column, 3.7e24 cm^-2 (Barnes+ 2019 Sec 3.1) = 8.3e4. The bars are drawn flat, at the cloud-averaged limit. A contour-by-contour treatment would tilt them upward to the right, since the area shrinks at higher contour — that is why Pokhrel's clouds are curves, not points. So flat is the conservative rendering at the high-column end.
Cross-check on which of the two is denser (Tang+ 2021, MNRAS 505, 2392, Fig. 10): their AzTEC+Herschel dust-SED fit gives one uniform N(H2) map of the whole CMZ at 7", so the Brick and Cloud E/F are measured the same way. Reading that map in matched R = 2.4 pc apertures: mean lg N(H2) = 22.90 for the Brick, 23.03 for cloud e and 23.01 for cloud f. Cloud E/F is the denser of the two, by ~0.13 dex (~40%). That is the same ordering the Walker+ 2018 masses give (5.1e3 vs 6.0e3 Msun/pc^2, 0.07 dex), and by a slightly larger margin — so the separation plotted here is the conservative one. Both clouds saturate Tang's 10^23.4 color scale over a few per cent of their area, so these are means, not peaks.
Two things I got wrong first, in case they come up: (i) I first took Sigma_gas from Barnes' intro prose, "~1e5 Msun, radii of ~1 pc". Multiplied together those give 3.2e4 Msun/pc^2, five to thirteen times too high, and put E/F to the right of most of Sgr B2. They are order-of-magnitude statements; Barnes never tabulate a cloud-scale mass or radius. Walker+ 2018 Table 1 is the number to use. (ii) I first set the E/F ceiling from Barnes' "the core region could be up to ~600 Msun ... capable of forming one or several high-mass stars (~25 per cent efficiency)". That 150 Msun is star formation potential, not stars that exist — those cores are starless — and treating it as already formed put the limit inside the Sgr B2 cloud.

Rebuild: python3 colloquium_assets/make_sf_relation_figure.py. Edit CMZ_CLOUDS at the top of that file and rerun; it prints this provenance and the dex-below-relation for every point. Local tracks are digitised from Pokhrel+ Fig. 2a (10 of 12 clouds; Orion-A and Mon R2 are black/grey and cannot be separated from the axes).

What is different about the CMZ?

JWST has observed the CMZ dust ridge

JWST's colors show ice

The Dust ridge: Cloud A (brick), C, D

All have different ice-to-dust ratios

Gramze+ (subm 2026)

Cloud D

Ice is everywhere

But it's different between clouds

3 kpc filament

Brick Head

Gramze+ (subm 2026)

Slope changes imply chemical differences between clouds within the CMZ

CO/Water ratio?

OCN⁻ / Ammonium salt abundance?

TBD: Ashby+ NIRSPEC 6927

Gramze+ (subm 2026)

We know stars are forming in the CMZ, but we've measured only a small portion

The JWST Galactic Center Treasury Survey (GO 10678) — NIRCam: running now

starformation.astro.ufl.edu/avm_images/jwst_gc_aladin.html
starformation.astro.ufl.edu/jwst-gc

Stellar populations: star formation history
Extinction maps: where is the gas?
YSO candidates

■ observed (12) ■ scheduled (62) ■ delayed (1) ■ awaiting a date (64) status 2026-09-12

…and the same survey with MIRI

starformation.astro.ufl.edu/avm_images/jwst_gc_aladin.html
starformation.astro.ufl.edu/jwst-gc

YSO candidates
PAH feedback maps

■ observed (12) ■ scheduled (62) ■ delayed (1) ■ awaiting a date (64) status 2026-09-12

Progress:

starformation.astro.ufl.edu/avm_images/jwst_gc_aladin.html
starformation.astro.ufl.edu/jwst-gc

38 of 139 pointings observed
F212N + F480M, opened at $\ell = +0.7$
a second block has opened at Sgr C

□ observed — NIRCam F212N/F480M mosaic of 2026-09-15, still growing

Built by colloquium_assets/make_cmz_treasury_layers.py, which is live in two ways and should be re-run shortly before the talk: the mosaics come from the growing HiPS at starformation.astro.ufl.edu/avm_images/jwst_gc_treasury_hips/ (NIRCam) and .../jwst_gc_treasury_miri_hips/ (MIRI F770W parallel), and the footprint colors come from the STScI visit status behind starformation.astro.ufl.edu/jwst-gc/monitor/. As of 2026-09-12: 139 pointings, 12 observed (7 Executed, 3 Archived, 2 Collecting), 62 Scheduled, 64 Flight Ready, 1 Skipped (GC_136 -- i.e. delayed, it will be rescheduled). Reduction lags the telescope: only 18% of the observed NIRCam area has made it into the mosaic so far, so the colored footprints run ahead of the picture. Caveat worth knowing but probably not worth saying: program 10678 has no offsets table yet, so these tiles sit on the raw assign_wcs frame with no measured tie to VIRAC2/Gaia. Positions are good to roughly a JWST pointing, not to a pixel. The survey runs east to west in descending visit number -- it opened at l = +0.70 and works down through +0.58, +0.51, +0.45 -- so the panel fills in from its left edge. Ten of the twelve observed pointings are the block visible at l = +0.64/+0.70; the other two (GC_138, GC_139) sit at b = +0.58, just above the top of this framing, so the slide shows 10 outlines while the bullet says 12.

…and the MIRI parallel, ~6′ to the east

starformation.astro.ufl.edu/avm_images/jwst_gc_aladin.html
starformation.astro.ufl.edu/jwst-gc

F770W: PAH emission, warm dust

□ observed — MIRI F770W mosaic of 2026-09-15, still growing

Built by colloquium_assets/make_cmz_treasury_layers.py, which is live in two ways and should be re-run shortly before the talk: the mosaics come from the growing HiPS at starformation.astro.ufl.edu/avm_images/jwst_gc_treasury_hips/ (NIRCam) and .../jwst_gc_treasury_miri_hips/ (MIRI F770W parallel), and the footprint colors come from the STScI visit status behind starformation.astro.ufl.edu/jwst-gc/monitor/. As of 2026-09-12: 139 pointings, 12 observed (7 Executed, 3 Archived, 2 Collecting), 62 Scheduled, 64 Flight Ready, 1 Skipped (GC_136 -- i.e. delayed, it will be rescheduled). Reduction lags the telescope: only 34% of the observed MIRI area has made it into the mosaic so far, so the colored footprints run ahead of the picture. Caveat worth knowing but probably not worth saying: program 10678 has no offsets table yet, so these tiles sit on the raw assign_wcs frame with no measured tie to VIRAC2/Gaia. Positions are good to roughly a JWST pointing, not to a pixel. The survey runs east to west in descending visit number -- it opened at l = +0.70 and works down through +0.58, +0.51, +0.45 -- so the panel fills in from its left edge. Ten of the twelve observed pointings are the block visible at l = +0.64/+0.70; the other two (GC_138, GC_139) sit at b = +0.58, just above the top of this framing, so the slide shows 10 outlines while the bullet says 12.

Roman covers the CMZ & more

Astrometry!
Variability!
$\rightarrow$ 3D tomography

□ Roman GC hourly-cadence □ Roman GBTDS PA 90.6° (spring window)

We have had access to position-velocity for a long time, but we still argue about position

Dani Lipman, Robin Tress, Mattia Sormani, et al — inner-Galaxy simulation, gas color-tagged by its true 3D position: face-on (top left), as we see it on the sky (bottom left), and in l–v (right)

Cloud Tomography with Stellar Proper Motions

Stellar orbital motion is measurable w/JWST + Roman

Clouds block and redden stars behind and within

The measurable stellar density with Roman is enough to achieve ~10 pc resolution

Interactive version; orbits after Nieuwmunster+ 2024; method after Vasini+ 2026

Cloud location inversion: ~10 pc resolution along the line of sight

~10⁴ stars → ~10 pc accuracy.

Observed density ~500 stars pc⁻² at m_K<19 gives ~10–20 pc

Roman GO 19008; Vasini+ 2026

The GC is the laboratory where cosmologically common conditions are observed at planetary-system forming scales.

STOP HERE

The cometary hypothesis: the molecules for life are built in the cloud

Complex organics form on icy grains in cold molecular clouds

…and survive the collapse into the protoplanetary disc

…where they are locked into comets and asteroids

…which deliver them, already made, to a young planet

Figure: V. M. Rivilla

…but the first link is not the same in every cloud

Ice is everywhere in the CMZ — but its composition is not the same from cloud to cloud.

Different CO/H₂O, different OCN⁻: a different mix of simple precursors goes into the disc.

If the CMZ is our local stand-in for cosmic-noon conditions, then the inventory delivered to planets there is not the Solar-neighbourhood one.

Gramze+ (subm 2026)

Figure: V. M. Rivilla

Why Roman: proper motions give us the third dimension

We see the CMZ in projection.

Orbital models disagree about what is in front of what — and the gas kinematics alone cannot settle it.

Roman GO 19008 (Ginsburg+) — CMZ FUZ

Extinction rises with distance — except where clouds intervene

A_V to each star is a running total along the line of sight
A cloud imprints a step in A_V at its distance
Map enough stars and you can invert A_V → d

Roman GO 19008 (Ginsburg+) — CMZ FUZ

Feasible in the first two seasons, with archival first epochs

Roman – HST (Paα): ~0.1 mas/yr, Δt ~ 19 yr
Roman – GALACTICNUCLEUS: ~0.3 mas/yr, Δt ~ 5–12 yr
Roman – VVV: ~0.5 mas/yr, Δt ~ 17 yr
Roman – JWST (GO 10678): ~0.5–1 mas/yr, Δt ~ 2–5 yr
Archival epochs give coarse (~50 pc) tomography immediately; Roman-internal PMs reach the ~10 pc goal by season two

Roman GO 19008 (Ginsburg+) — CMZ FUZ

The method exists, and it works on mock data

The nuclear stellar disc rotates: near side toward positive longitude, far side toward negative — a ~5 mas/yr difference against ~1 mas/yr scatter. A proper motion therefore places a star on the near or the far side, and its extinction measures the dust in front of it.

Proper motions + extinctions → a non-parametric 3D dust map, with no orbital model assumed.

~100 stars per 10×10 pc beam → 20 pc resolution; ≳1000 → 4–8 pc. GALACTICNUCLEUS+HST already gives ~25 stars pc⁻², VIRAC2 ~50, JWST GO 10678 ~150 after one epoch and ~1500 after two.

Vasini, Sormani, Sanders, Ginsburg, Battersby+ 2026 (arXiv:2609.10782)

This is the companion to the previous three slides: they make the case, this is the machinery, and it is on the arXiv as of two days ago
Big panel is their Fig. 7: a hydrodynamical-simulation CMZ (left) and what the method recovers from mock stars (right), on the longitude–line-of-sight plane. The ring, its near/far asymmetry and the individual clumps all come back
Inset is their Fig. 1: the rotation curve of the NSD in proper motion. Near side +2.5, far side −2.5 mas/yr — that split IS the distance information
The requirement is modest: ~1–10 stars pc⁻². Existing surveys already clear it; the paper names Roman GO 19008 among the programmes that will deliver it
Honest caveat: this is a proof of concept on mocks. Before real data they still need the nuclear stellar cluster and the bar added, velocity errors, and a real selection function

A hundred stars, a hundred orbits — how proper motions locate a cloud

Interactive version: assets/cmz/cmz_orbit_stars.html — model after Nieuwmunster+ 2024; method after Vasini+ 2026