Overview¶
Galaxy SEDs encode stellar mass assembly, chemical enrichment, and dust processing. Tengri recovers these from broadband photometry and spectra using one forward model spanning stellar populations through X-ray, all driven by shared parameters.
Built on JAX and DSPS. Inference backends (optimizers, samplers, variational inference) plug in as registrations. tengri.summary() prints the live count for every registry; new components register themselves.
Philosophy¶
The forward model is the artifact. Inference is a thin shell, so backend swaps are one-line changes. Physics lives in components and instruments in observation, so adding a dust law or sampler is a registration rather than a rewrite.
What’s modular¶
Layer |
What you can swap |
|---|---|
Stellar SSP |
BC03, BPASS, FSPS, ProGeny: any HDF5 file in the DSPS schema |
SFH |
parametric (about 15 families), non-parametric (Leja+ continuity, Dirichlet), and stochastic (IFT correlated fields with PSD-governed burstiness) |
Dust attenuation |
Calzetti, Cardelli, Charlot & Fall two-component, Salim, Kriek & Conroy, others |
Dust emission |
Draine & Li, Dale, THEMIS |
Nebular |
|
AGN |
accretion disc, SKIRTOR torus, BLR/NLR through Cue, optional X-ray and radio |
IGM |
Madau, Inoue |
Inference |
optimizers from |
tengri.summary() prints the live count for every registry, and
new components register themselves; nothing in this table is updated
by hand.
Architecture¶
The forward model is layered. Each layer does one thing and talks to the layer below it through one small Protocol.
┌─────────────────────────────────────────────────────────────┐
│ Inference optax · BlackJAX · NIFTy.re · Population │
│ ↳ asks ForwardModel for log p(data | params) │
└─────────────────────────────────────────────────────────────┘
│ knows nothing about physics
▼
┌─────────────────────────────────────────────────────────────┐
│ Likelihood Gaussian · StudentT · Composite │
│ ↳ scores Observation.predict(params) against data │
└─────────────────────────────────────────────────────────────┘
▲
│ prediction dict
┌─────────────────────────────────────────────────────────────┐
│ ForwardModel populations (SEDModel + SpatialModel) │
│ owns Parameters, threads ForwardState │
│ ↳ .predict(params) → dict ──────────► Observation │
└─────────────────────────────────────────────────────────────┘
▲
│ threads state through components
┌─────────────────────────────────────────────────────────────┐
│ Components Stellar · SFH · Dust attn. · Dust IR · │
│ Nebular · AGN · IGM · Radio · X-ray │
│ each: parameters, precompute, apply(state) │
└─────────────────────────────────────────────────────────────┘
▲
│
┌─────────────────────────────────────────────────────────────┐
│ Observation Photometry · Spectroscopy · │
│ FiberSpectroscopy · Imaging · Joint │
│ ↳ reads ForwardState, returns prediction dict │
└─────────────────────────────────────────────────────────────┘
The rule the whole stack follows is physics in components and instruments in observation. A fiber aperture correction lives in FiberSpectroscopyObservation, not in any component, and a per-age spatial profile lives in a SpatialComponent, not in ImagingObservation. Keeping that line straight allows the codebase to stay legible after eight or nine layers of physics stack into a single fit.
Components¶
Each component declares parameters, reads from upstream, publishes to downstream, and evaluates on a wavelength grid. A new dust law or AGN library is one file; the registry and JIT plumbing auto-discover it.
Energy balance is enforced by contract: Stellar emits L_ν, Dust attenuation publishes L_absorbed, Dust IR emission consumes it. Nebular, AGN, IGM, radio, and X-ray layer on top. Spatial components (Sérsic, exponential, bulge+disk) follow the same pattern.
Sub-models¶
An SEDModel is components plus a validator that closes the reads/publishes graph. A SpatialModel does the same spatially. Hierarchical SFHs, per-age morphology, and stellar atmospheres are components or sub-models, not special cases.
ForwardModel¶
What inference talks to. Owns Parameters, holds populations (sub-model + optional spatial model), wires to Observation, exposes .predict(params). Multi-population from day one: AGN point source, Sérsic bulge, exponential disc are three populations with namespaced parameters.
Observation¶
Instruments live here. Photometry integrates through filter curves. Spectroscopy resamples plus LSF. FiberSpectroscopy corrects aperture via spatial profile integration, making joint photo plus fiber fits one-liners rather than hacks in components. Imaging and joint observations follow the same pattern.
Likelihood and inference¶
Likelihood (Gaussian, Student-t, composite) scores predictions against data. Inference asks for \(\log p(\text{data} \mid \text{params})\) and nothing else. Backends plug in unchanged.
Why JAX¶
Automatic differentiation: \(\partial \log p / \partial\theta\) for free at the cost of one extra forward pass. Stays correct under interpolation and emulators, where finite differences fail.
JIT compilation makes the forward model competitive with C. vmap vectorizes across galaxy samples. One source runs on CPU, GPU, TPU.
JAX’s discipline (pure functions, declared shapes, no mutation, no Python if on traced values) costs upfront. We absorb that cost in the library.