slabx

A Python reimplementation of SLAB, the dense-gas dispersion model of Ermak (1990, UCRL-MA-105607), verified against the original Fortran and evaluated against 38 field and wind-tunnel trials.

The implementation follows the published model formulation and the user's manual, in a program structure of its own; individual results were then compared against the original Fortran, which was also consulted for implementation details the manual leaves open. It is therefore not a clean-room implementation, and the comparison points are recorded in the source as line references so that the checking is visible.

from slabx.core.plume import run_dispersion
from slabx.core.source import EvaporatingPool
from slabx.post.concentration import concentration_field
from slabx.submodels.atmosphere import Atmosphere
from slabx.thermo.base import LegacyThermo, Substance, water_backend

LNG = Substance(name="LNG", mw=0.016043, cp_vapour=2238.0, cp_liquid=3348.5,
                dh_vap=509900.0, T_boil=111.7, rho_liquid=424.1)

atm = Atmosphere(u_ref=1.94, z_ref=3.0, T=290.0, rh=50.0, z0=2e-4,
                 stability="E")

# The pool area is the spill divided by the evaporation flux, not the
# footprint quoted in the manual, and the manual's worked example uses
# demonstration weather rather than the trial's. Both matter: the manual's
# 657 m2 with T = 306 K and 4.6 % humidity gives 386 m instead of 449.
src = EvaporatingPool(substance=LNG, rate=116.93, area=116.93 / 0.167,
                      duration=107.0)

traj, used = run_dispersion(src, atm, LegacyThermo(LNG), water_backend(),
                            x_max=2000.0, n_puff_steps=40)
field = concentration_field(traj, atm, z=1.0, t_avg=80.0, t_release=107.0)

print(f"{field.distance_to(0.05):.0f} m to LFL")     # 449; measured 455

slabx.validation.lng_pools sets all of this from the trial record, which is the path the paper's numbers come from -- but it needs the observations, which are not distributed here. See below.

Install

pip install -e .            # numpy, scipy
pip install CoolProp        # optional: real-fluid properties

Python 3.10+. QUICKSTART.md has a runnable scenario template and a troubleshooting table.

What it does

Four source types — evaporating pool, horizontal jet, vertical jet with plume rise, instantaneous release — integrated as a shallow-layer model with the plume-to-puff transition of the original.

Beyond the original, all off by default so the reproduction is never disturbed:

Thirty-eight coefficients are exposed as a frozen dataclass with presets, so sensitivity studies do not need edits to the source.

How far it has been checked

Against the original Fortran. The five decks of the manual agree to 0.05–1.48 % on every cloud variable, at the level of the original's own discretisation error. A random differential test over 108 decks — four source types, five materials, the full stability range — puts the median difference at 0.09 % on temperature and 0.67 % on concentration.

The comparison is against Ermak's Fortran itself, compiled from source. Early work used a JavaScript transcription and misattributed five of six differences to SLAB as a result; docs/02_VALIDATION_REFERENCE.md §2.5 records what that cost and how it was found.

Against measurements. Ten LNG pool trials (Burro, Coyote, Maplin Sands), three Desert Tortoise and three FLADIS two-phase ammonia jets, fifteen Thorney Island instantaneous releases, two Prairie Grass passive-dispersion trials, and three wind-tunnel trials. Nineteen further SMEDIS trials are excluded with a recorded reason for each.

On the ten LNG trials the model passes all five Chang–Hanna criteria (FAC2 0.90, MG 1.25, NMSE 0.12).

Where it fails. LIMITATIONS.md is the honest list: a known runaway for dense, low-momentum releases; a stability response that four independent datasets say is too strong; a hard-coded reference height that breaks Froude similarity; and two published datasets for the same trials that disagree by a factor of two and reverse the recommended surface roughness.

VALIDATION.md has the numbers and the commands that reproduce them.

Tests

pytest -q
python3 examples/validate_burro.py     # field comparison, pre-registered
python3 examples/diagnose_burro.py     # how the hypotheses were excluded
python3 -c "from slabx.scope import describe_scope; print(describe_scope())"

Tests that need the original SLAB source, an optional dependency, or a third-party observation are skipped when the resource is absent, so the counts depend on what you have installed and obtained. Nothing fails for want of them.

The one xfailed is deliberate: a reproducible failure kept with its input deck rather than hidden.

Comparing against the original needs its source, which is not redistributable here; see golden/fortran/README.md. Those tests skip without it and everything else runs.

Documentation

Citing

Cite the version you ran -- the version DOI fixes the files, the concept DOI resolves to the latest:

Lee, U. (2026). slabx: a Python reimplementation of the SLAB dense-gas dispersion model (v1.0.2). Zenodo. Concept DOI 10.5281/zenodo.22023625.

CITATION.cff carries the same in machine-readable form.

The model itself is Ermak's:

Ermak, D.L. (1990) User's Manual for SLAB: An Atmospheric Dispersion Model for Denser-than-Air Releases. UCRL-MA-105607, Lawrence Livermore National Laboratory.

Licence

MIT, for this implementation. THIRD_PARTY_NOTICES.md records the third-party works relied on, none of which are included here. The original SLAB source is not included and carries its own non-commercial terms; see LICENSE and golden/fortran/README.md.