BIND — Max E. Lee

Animation of flow matching: a gas field emerging from pure noise into a cluster-mass halo — Watch the generation itself: BIND integrates a learned velocity field from pure noise (t = 0) to the finished gas field (t = 1) of a cluster-mass halo — about a second of GPU time.

BIND (Baryonic INpainting with Deep learning) is a conditional flow-matching model trained on the 1024 paired hydrodynamical and dark-matter-only simulations of the CAMELS SB35 suite. Given a dark-matter-only halo, it generates the corresponding dark matter, gas, and stellar mass fields — conditioned on the full 35-dimensional ΛCDM and IllustrisTNG galaxy formation parameter space.

Applied halo-by-halo to full N-body volumes and ray-traced, BIND produces convergence, optical depth, and Compton-y maps whose statistics match full hydrodynamical simulations to within the precision of upcoming surveys. The trained models, generated halos, and map suites are released as open-source tools.

A wide slice of a 50 Mpc/h simulation volume: BIND-generated gas pasted onto the halos of the dark-matter cosmic web — BIND gas pasted halo-by-halo into the dark-matter web of a 50 h⁻¹ Mpc volume.

One input, seven fields

From a single dark-matter-only conditioning field, BIND generates every baryonic channel at once — masses and thermodynamics — for the same halo in one draw:

Dark-matter-only input field — DMO input

Generated dark matter field — dark matter

Generated temperature field — temperature

The papers

BIND (Baryonic INpainting with Deep learning): A Field-level Emulator for Galaxy Groups and Clusters — the model: training, validation, and halo-level performance. Download PDF
BINDing the lightcone: A suite of astrophysical ray-traced weak lensing and SZ maps — the maps: 1000 pseudo-independent realizations across the IllustrisTNG galaxy formation prior. Download PDF

Turning the astrophysical dials

BIND is conditioned on every IllustrisTNG galaxy formation parameter, so the same halo can be regenerated under different astrophysics. Here the same dark-matter-only halo is painted while one feedback parameter sweeps across the full SB35 prior (cyan tick) with everything else — including the initial noise — held fixed:

Animation of the same halo's gas field as the supernova wind energy parameter sweeps across its prior, with a position indicator below — A_SN1 · galactic wind energy · prior low → high

Animation of the same halo's gas field as the AGN radio feedback parameter sweeps across its prior, with a position indicator below — A_AGN1 · AGN radio feedback · prior low → high

And because BIND is generative rather than deterministic, every draw is a new plausible realization — same halo, same parameters, four different draws:

Four side-by-side BIND draws of the same halo with identical parameters, each showing different small-scale structure — one halo · one parameter vector · four independent draws

Does it get the physics right?

Start with the eye test — the same merging cluster, once from the full IllustrisTNG-300 simulation and once drawn by BIND from the dark-matter-only input alone, on a shared color scale:

Gas field of a merging cluster from the IllustrisTNG-300 hydrodynamical simulation — IllustrisTNG-300 truth

Gas field of the same merging cluster generated by BIND from the dark-matter-only input — BIND draw from the DMO input

Beyond looking right, the generated halos reproduce the target statistics. Pasted back into an N-body volume, BIND recovers the projected matter power-spectrum suppression to the accuracy ceiling set by pasting in the true hydrodynamical halos themselves:

Power spectrum ratio P(k) over P DMO of k, showing BIND tracking the truth and the hydro-replaced ceiling — Projected matter power-spectrum suppression: BIND (red) against the IllustrisTNG truth (black) and the hydro-replacement ceiling (dashed).

Try it live

The explorer lets you tune cosmological and astrophysical parameters with sliders and watch BIND generate halo projections in real time, running inference on a live GPU.

⚠️Note: the app runs on a cloud GPU and may take up to a minute to wake on first visit.

Launch the BIND Explorer ↗