Concepts and known approximations

lazuli_exo is a comparison tool, not a forecasting tool. It exists to let you ask “how do these atmospheres, targets, or transit counts compare under one consistent set of assumptions?”, and it is explicit about which assumptions those are.

Warning

Every scenario is a toy H2/He atmosphere with idealized noise. No number produced by this package is a calibrated detection significance or an observing-time forecast. Treat outputs as relative and qualitative.

The measurement in one page

If transmission spectroscopy is new to you, this is the whole idea.

When a planet crosses its star, the star dims by the fraction of its disk the planet covers, the transit depth \(\delta = (R_p/R_s)^2\). At a wavelength where the planet’s atmosphere absorbs, the planet looks slightly larger, so the transit is slightly deeper. The size of that extra depth is set by the atmospheric scale height \(H = k T_\mathrm{eq} / (\mu m_H g)\): one scale height of extra radius adds about \(2 R_p H / R_s^2\) to the depth. For a hot Jupiter that is a few hundred parts per million (ppm); for a temperate sub-Neptune, tens of ppm. The transmission spectrum is the transit depth measured as a function of wavelength, and its bumps are the molecular bands (water near 1.4 µm, sodium and potassium in the optical) and the continuum slope of a haze.

lazuli_exo computes two things for such a measurement with Lazuli. The model spectrum, binned to a resolving power R (the number of wavelength bins per octave divided by \(\ln 2\); R = 100 gives about 145 bins over the band). And the 1-sigma uncertainty on the depth in each bin for n_transits co-added transits, from the host star’s brightness, the transit duration and the instrument’s photon and read noise. From those two arrays it derives a single number, the template S/N, which says how far the model spectrum sits from a flat line in units of the error bars. It is a comparison statistic between planets, scenarios and observing strategies; The template statistic explains why it is not a detection significance.

A few terms used throughout:

depth, depth_err: The binned model transit depth (a fraction; multiply by \(10^6\) for ppm) and its 1-sigma uncertainty per bin.
R: Resolving power of the binned grid, \(\lambda / \Delta\lambda\). Coarser bins (smaller R) have smaller error bars and resolve less.
n_transits: Transits co-added; error bars shrink as \(1/\sqrt{N}\) with no floor.
feature_amplitude_ppm: Peak-to-trough of the binned model over 0.40 to 1.70 µm, in ppm.
template_snr_diagonal: The template S/N described above.
scenario: A preset of the atmosphere parameters: clear, cloudy, hazy, high_metallicity or flat (Spectra and scenarios).

The two-stage architecture

The heavy physics runs once, offline, in environments that never have to coexist:

inputs (POSEIDON opacities + fastchem grid, PHOENIX stellar grid)
      |
      v  poseidon conda env
atmo/zeta_grid.npz     star/phoenix_grid.npz     star/<planet>.npz
(opacity shape on a (T, [M/H], C/O) grid)             |
      |                                               v  py313 env + slicersim
      |                                         noise/<planet>.npz
      |                                         (single-transit depth_err(lambda))
      |                                               v  py313 env + exotic-ld
      |                                         ld/<planet>.npz
      v                                         (quadratic limb darkening)
runtime (src/lazuli_exo/) = numpy + these tables, nothing heavier

precompute/tables/models/<planet>__<scenario>.npz   POSEIDON validation data

POSEIDON and slicersim are never installed together. The runtime imports neither. This is what lets the package be deployed behind a web app without shipping tens of GB of opacity inputs.

Known approximations

These are the limits that matter. Each is real, each is knowable, and each is recorded somewhere you can check.

Isothermal equilibrium chemistry: Every spectrum uses a single \(T = T_\mathrm{eq}\) isotherm, not a real pressure-temperature profile, with equilibrium chemistry throughout (the scenario presets fix C/O = 0.55; simulate() lets you vary it).
Surrogate atmosphere: The spectrum is an interpolation of a POSEIDON-derived opacity grid plus analytic scale-height, column, deck and haze terms, not a live radiative-transfer run. Inside its validated box it matches POSEIDON to well under the single-transit noise; see The parametric model for the box and the gates.
Cold-planet temperature floor: The fastchem equilibrium-chemistry grid has a hard floor at 300 K. K2-18 b (\(T_\mathrm{eq} \approx 284\) K), LHS 1140 b (≈ 226 K), and TRAPPIST-1 e (≈ 250 K) use 300 K chemistry, while the scale height uses the true \(T_\mathrm{eq}\). Recorded on every observation as meta["atmosphere"]["T_clamped"] (and T_chem_K).
\(\sqrt{N}\) S/N scaling: depth_err scales as \(1/\sqrt{N_\mathrm{transits}}\) from a single precomputed single-transit uncertainty. There is no per-transit systematics term and no red-noise floor, so predicted precision improves without limit as transits are added. This is the single largest unrealism in the package. See The noise model.
Equal out-of-transit baseline: The noise model assumes an out-of-transit baseline exposure equal to the in-transit one, i.e. \(\sigma_\mathrm{depth} = \sqrt{1/\mathrm{SNR}_\mathrm{in}^2 + 1/\mathrm{SNR}_\mathrm{out}^2}\) with \(\mathrm{SNR}_\mathrm{in} = \mathrm{SNR}_\mathrm{out}\), not an optimized or telescope-time-constrained baseline.
Template S/N, not detection significance: detect_features() returns a matched-template amplitude S/N against the best-fit flat line under a diagonal covariance. There is no defined null/alternative likelihood ratio, no nuisance model, no covariance, and no false-positive calibration. See The template statistic.
Residual detector saturation on bright hosts: For HD 189733 b, slicersim’s detector ramp cannot be fully de-saturated even at the minimal MACC configuration (2 groups × 1 frame). The residual pixel count is recorded as meta["noise"]["nsaturated_detpx"] and observe() raises a UserWarning. Its uncertainties are unvalidated: saturation can bias extraction weights, nonlinearity, and wavelength dependence in either direction. Check the flag rather than the name; the list changes when the tables are regenerated.
Light-curve geometry: Quadratically limb-darkened transits with geometry from a frozen NASA Exoplanet Archive registry. Planets outside the registry, or without a limb-darkening table, fall back to a \(b = 0\) trapezoid; meta["shape"] records which. See Transit light curves.

The caveats live in the data

Two of the approximations above are flagged per table, so code can branch on them rather than relying on someone having read this page:

obs = lazuli_exo.observe("K2-18 b", scenario="clear")
obs.meta["atmosphere"]["T_clamped"]       # True: chemistry grid floor hit

obs = lazuli_exo.observe("HD 189733 b")   # also raises a UserWarning
obs.meta["noise"]["fully_desaturated"]    # False
obs.meta["noise"]["nsaturated_detpx"]     # residual saturated pixels

The upgrade paths

The two most consequential limitations have concrete, non-speculative fixes that fit the existing table schema without API changes:

Molecule detections. Replace the band-restricted template norms with a likelihood ratio against per-molecule leave-one-out model tables. The schema already supports additional scenario names per planet.
Realistic transit counts. Add a per-transit systematics or red-noise term to the noise tables, which would break the pure \(\sqrt{N}\) scaling and stop transit counts from being optimistic without bound.