Atmospheric dispersion of wildfire smoke plumes: HYSPLIT, FLEXPART, SILAM, NAME, CALPUFF — Lagrangian vs Eulerian

April 12, 2026

Posted in Blog

Atmospheric dispersion of wildfire smoke plumes: HYSPLIT, FLEXPART, SILAM, NAME, CALPUFF — Lagrangian vs Eulerian

Lagrangian vs Eulerian paradigm — the fundamental difference

The classical Eulerian model solves the advection-diffusion-reaction equation for the concentration C(x,y,z,t) of a pollutant on a fixed spatial grid: ∂C/∂t + ∇·(uC) – ∇·(K∇C) = E – L, where u is the wind field, K is the turbulent diffusion tensor, E is the emission source term, and L represents sinks (dry and wet deposition, chemistry). The grid is typically 1-12 km for regional models and 25-50 km for global ones. The result is a concentration field on the grid at every time step.

The Lagrangian model simulates transport through an ensemble of computational particles (effective “carriers” of pollutant mass). Each particle moves in the wind field u + u’, where u’ is a stochastic turbulent perturbation (typically via a Langevin equation with adaptive Lagrangian time scales). Dry and wet deposition are computed as a reduction of particle mass. Concentration is reconstructed post-hoc through a kernel-density approach on the desired output grid (typically 1-3 km for wildfire smoke).

The conceptual advantage of the Lagrangian approach is the absence of artificial diffusion from numerical schemes, optimal behaviour for narrow sources and short trajectories (point sources, the first 24-72 hours after emission). The conceptual advantage of the Eulerian approach is the natural inclusion of chemical reactions (with native ODE-grid resolution per step), straightforward incorporation of secondary aerosols, and better behaviour for long (5-30 days) and continental-scale transport.

HYSPLIT — the NOAA flagship since the 1980s

HYSPLIT (HYbrid Single-Particle Lagrangian Integrated Trajectory) is the most widely used Lagrangian model in the world. Developed by NOAA Air Resources Laboratory in the 1980s, it is described in the canonical paper Stein et al. 2015 (BAMS 96:2059-2077). HYSPLIT supports two modes: pure trajectory mode (tracking an air parcel without dispersion) and dispersion mode (a full particle or puff tracker).

HYSPLIT uses meteorological input from NCEP/NCAR Reanalysis, GFS, NAM, NARR, RAP, HRRR (for the US); ECMWF ERA5 and IFS (globally); and WRF output (custom). The time step ranges from 1 minute to 1 hour adaptively. It supports dry and wet deposition, distance-dependent decay (for radionuclides), and band-dependent photochemical loss. It does not include full gas-phase chemistry.

HYSPLIT is used by the US National Weather Service for operational smoke forecasting (through integration with the BlueSky framework from USFS), by many national meteorological services for radionuclide incidents (Chornobyl 1986 — many post-event analyses, Fukushima 2011), and in atmospheric transport research. Documented validation: Su et al. 2015 (Atmos. Chem. Phys.) on Cesium-137 transport after Fukushima.

FLEXPART — the European Lagrangian benchmark

FLEXPART (FLEXible PARTicle dispersion model) is the second principal Lagrangian model, originating at the Norwegian Institute for Air Research (NILU) and further developed at Vienna University of Technology (TUW). The canonical publication is Stohl, Forster, Frank, Seibert, Wotawa 2005 (Atmos. Chem. Phys. 5:2461-2474). The current FLEXPART 10.4 version is described in Pisso et al. 2019 (Geosci. Model Dev. 12:4955-4997).

FLEXPART has several advantages over HYSPLIT for wildfire smoke: more efficient handling of large particle ensembles (10-100 million particles), more realistic parameterisation of convective transport (the Emanuel convection scheme), and built-in support for backward simulation (inverse modelling — critical for source attribution). FLEXPART works with ECMWF ERA5/IFS, GFS, and WRF output. One canonical use case is Stohl et al. 2012 (ACP) on the 2007 Caribou Hills fire in Alaska and its global black-carbon transport.

Backward FLEXPART is a separate class of application. Instead of releasing particles from a known source forward in time, particles are released from a known receptor backward. The result is a sensitivity field showing from which regions and at which times significant contributions reached the receptor. This is the basis for inverse modelling of sources — a technique that allows estimation of unknown emission rates from observations.

SILAM — the Finnish multi-mode model

SILAM (System for Integrated modeLling of Atmospheric coMposition) from the Finnish Meteorological Institute (FMI), Sofiev group, is the third major operational atmospheric chemistry and dispersion model, with a unique architecture: it supports both Lagrangian and Eulerian modes within a single codebase. The canonical publication is Sofiev et al. 2015 (Geosci. Model Dev. 8:3497-3522).

SILAM is used operationally in CAMS (Copernicus Atmosphere Monitoring Service) as one of nine member models in the regional ensemble over Europe. Its unique strength is IS4FIRES, an integrated system for emission estimation from fires with assimilation of satellite FRP data, which feeds the operational SILAM-Fire module. Sofiev, Vankevich, Lotjonen et al. 2009 (Atmos. Chem. Phys. 9:6833-6847) describes the IS4FIRES methodology.

The partnership with FMI is strategic for WildFiresUA: SILAM produces operational smoke plume forecasts for Europe with 1 km regional downscaling over Ukraine, which we use as a reference for validating our WRF+FLEXPART stack. Joint work with Mikhail Sofiev and his group is part of the peer-reviewed publication pipeline.

NAME — the UK Met Office stack

NAME (Numerical Atmospheric-dispersion Modelling Environment) is the UK Met Office stack that replaced the earlier MESOS system in the 2000s. The canonical publication is Jones, Thomson, Hort, Devenish 2007 (Air Pollution Modelling and its Application XVII). NAME is a Lagrangian particle-tracking code with its own parameterisation of turbulence in the boundary layer and aloft.

NAME is used by the UK Met Office across the full application spectrum: volcanic ash plumes (Eyjafjallajökull 2010 — the best-known example), radionuclide incidents (Chornobyl 1986, Fukushima 2011, routine Sellafield releases), industrial-emergency releases (Buncefield 2005), and many wildfire-related applications. One of NAME’s key strengths is rigorous verification across short and long transport ranges through inclusion in WMO Global Atmospheric Watch (GAW) verification exercises.

NAME uses the Met Office UM (Unified Model) NWP as input fields — making it tightly integrated with the UK national meteorological infrastructure. A similar philosophy of tight NWP+dispersion integration appears in the French MOCAGE (Météo-France) and German EURAD (Universität zu Köln).

CALPUFF — the Gaussian puff regulatory standard

CALPUFF is a Lagrangian puff model based on the Gaussian puff approximation for resolving the concentration field. Developed by Sigma Research Corporation in the 1980s-1990s, the canonical documentation is Scire, Strimaitis, Yamartino 2000 (CALPUFF documentation for US EPA). CALPUFF was for a long time the US EPA-approved model for regulatory long-range transport modelling (over 50 km from source) for Class I airshed exposures in the National Park Service and US Forest Service context.

CALPUFF occupies a unique niche: an official model for radiological scenarios in the IAEA context (although HYSPLIT and JRODOS are more popular in the European context); often used for long-range transport of industrial emissions and for assessing the impact of oil and gas facilities on protected areas. For wildfire smoke, CALPUFF is less popular than HYSPLIT and FLEXPART because it lacks native integration with GFAS-like emission inventories.

WildFiresUA uses CALPUFF as the principal model for radiological scenarios at nuclear power plants (Zaporizhzhia NPP, Khmelnytskyi NPP, South Ukrainian NPP) — a separate branch of our stack oriented toward civil protection in the event of a radiation incident.

WRF-Chem — the flagship Eulerian air-quality model

WRF-Chem is a coupled meteorology-chemistry model that links the mesoscale NWP model WRF with full gas-phase photochemical mechanisms (CB05, RACM, MOZART) and aerosol-phase modules (GOCART, MAM, MOSAIC). The canonical publication is Grell et al. 2005 (Atmos. Environ. 39:6957-6975); the current version is documented in Skamarock et al. 2019 (NCAR Tech. Note).

WRF-Chem is typically run on grids of 4-12 km (regional) down to 1-2 km (local). It captures full chemical evolution within the plume — formation of ozone, secondary ammonium nitrate, secondary organic aerosols — which is unattainable for Lagrangian models without post-processing chemical modules. A canonical wildfire-smoke application is Pfister et al. 2014 (Atmos. Chem. Phys.) on the 2008 California fire season.

The drawback is computational cost. A 24-hour forecast on a 4 km grid for a 1000×1000 km domain with 30+ aerosol species requires 4-10 hours on 200-400 HPC cores. This makes WRF-Chem practical for research-oriented post-event analyses and national-scale operational use by state meteorological centres, but expensive for regional agencies.

CMAQ — the US EPA standard

CMAQ (Community Multiscale Air Quality model) is the flagship US EPA Eulerian model, with a history dating to the 1990s. The canonical publication is Byun & Schere 2006 (Appl. Mech. Rev. 59:51-77). CMAQ uses MM5 or WRF as a driver for meteorological fields and includes full photochemical mechanisms (CB6, SAPRC) and an aerosol model (AERO6). Grids are 4-36 km (regional) and 1-4 km (urban).

CMAQ is the operational standard for US EPA NAAQS (National Ambient Air Quality Standards) compliance modelling: state agencies use CMAQ for long-term (annual) simulations on a 12 km grid to evaluate control strategies. For wildfire smoke, CMAQ has a dedicated PMc/PMf module that includes primary emissions from GFED4/FINN/QFED.

The European analogue of CMAQ is EMEP MSC-W (Cooperative Programme for Monitoring and Evaluation of the Long-range Transmission of Air Pollutants in Europe). EMEP MSC-W is the standard for CLRTAP (Convention on Long-range Transboundary Air Pollution) reporting. The canonical description is Simpson et al. 2012 (Atmos. Chem. Phys. 12:7825-7865).

Lagrangian vs Eulerian intercomparison — Vira 2022 and others

The most recent large-scale systematic comparison of Lagrangian (FLEXPART, SILAM-Lagrangian) and Eulerian (SILAM-Eulerian, EMEP MSC-W) approaches for wildfire smoke transport is Vira et al. 2022 (Geosci. Model Dev. 15:3963-3982). The paper uses SILAM in both modes for standard benchmark cases and shows: for short trajectories (12-48 hours), Lagrangian gives 20-40% better agreement with station observations; for long trajectories (5-15 days), Eulerian catches up and in some cases performs better by virtue of including chemical evolution.

An analogous study for radionuclide transport is Sato et al. 2018 (Atmos. Chem. Phys.) on the Fukushima case, with six models (HYSPLIT, FLEXPART, SILAM, MATCH, JRODOS, NAME). Inter-model spread is a factor of 2-5 in peak concentrations and a factor of 1.5-3 in peak timing. This shows that no single model “wins” unconditionally — the choice depends on task type and available computational resources.

Integrated operational practice across regional centres

United States. NOAA runs HYSPLIT operationally for smoke forecasting through integration with the USFS BlueSky framework and HMS (Hazard Mapping System). NCAR runs WRF-Chem for research applications and retrospectives. EPA runs CMAQ for regulatory compliance.

Europe. CAMS (Copernicus Atmosphere Monitoring Service) runs an ensemble of 9 models: SILAM (FMI), MOCAGE (Météo-France), CHIMERE (CNRS-IPSL-INERIS), DEHM (Aarhus University), EMEP MSC-W (Norwegian Met Inst), EURAD-IM (FZJ Jülich), GEM-AQ (Environment Canada), LOTOS-EUROS (TNO), MATCH (SMHI). Ensemble averaging gives a more robust forecast than any single model. The UK Met Office runs NAME separately.

Canada. Environment and Climate Change Canada runs GEM-MACH (Global Environmental Multi-scale model with Modelling Air quality and CHemistry) operationally; for smoke specifically, the FireWork module integrates with CWFIS (Canadian Wildland Fire Information System).

Australia. The Bureau of Meteorology runs AAQFS (Australian Air Quality Forecasting System) operationally, based on the CSIRO CTM; for smoke specifically, integration with CSIRO C-CTM and Sydney University AERMOD-style short-range models.

Regional comparison — operational stacks

Country / region	Operational smoke model	Type	Key partner
United States	HYSPLIT + BlueSky	Lagrangian	NOAA ARL + USFS AirFire
Canada	GEM-MACH + FireWork	Eulerian	ECCC
Europe (CAMS)	9-model ensemble	Mix	ECMWF
UK	NAME	Lagrangian	UK Met Office
Finland / Scandinavia	SILAM + IS4FIRES	Hybrid	FMI
Australia	AAQFS + C-CTM	Eulerian	CSIRO + BoM
Ukraine	WRF + FLEXPART (CALPUFF for radiological)	Lagrangian	WildFiresUA

Chemical evolution within the plume — beyond simple advection

Wildfire smoke does not remain stable within the plume. In the first 3-12 hours after emission, active heterogeneous oxidative chemistry occurs: NMVOCs are oxidised to OVOCs, OVOCs are further oxidised to more oxidised species and condense into secondary organic aerosol (SOA). NOx is converted to HNO3 (daytime) or N2O5 (nighttime). Black carbon interacts with OH radicals, acquires a hydrophilic shell, and increases its effective optical cross-section.

Canonical studies of plume evolution: Akherati et al. 2020 (J. Geophys. Res.: Atmospheres) on SOA formation in FIREX-AQ plumes; Decker et al. 2021 (J. Geophys. Res.: Atmospheres) on nighttime chemistry; Lim et al. 2020 (Atmos. Chem. Phys.) on optical analysis of BrC ageing.

Lagrangian models typically do not simulate this chemistry (with the exception of some FLEXPART-WRF-Chem extension modules). Eulerian models (CMAQ, WRF-Chem, EMEP MSC-W) do — this is their main advantage for long-range transport. For an operational 24-72 hour wildfire smoke forecast, the difference in ground-level concentration due to chemistry is typically 10-30%, which is significant for PM2.5 attribution to fire sources but not for the basic transport-pathway forecast.

Open problems and frontiers

Plume rise injection height. The height at which smoke reaches the atmosphere after emission critically affects subsequent transport. Ground-level injection produces short local transport; upper troposphere or stratosphere injection produces global transport. Standard parameterisations (Briggs 1975, Sofiev 2012) are validated but carry significant uncertainty for extreme cases. Sofiev et al. 2012 (Atmos. Chem. Phys.) proposed an improved parameterisation based on FRP. This remains a frontier for PyroCb cases.

Ensemble and probabilistic forecasting. A single deterministic forecast does not convey uncertainty. Probabilistic forecasts through ensemble NWP drivers and ensemble emission realisations are a frontier, partially implemented in CAMS but in need of refinement for wildfire smoke specifically. Lannelongue et al. 2022 (Atmos. Chem. Phys.) describes the ensemble approach for CAMS.

Machine learning as an emulator. An ML emulator of full Eulerian or Lagrangian models can deliver an operational forecast with physical-model quality in minutes instead of hours. Chen et al. 2023 (Mon. Weather Rev.) is an example of an ML emulator for air quality.

Multi-scale integration. Urban-plume interaction between large regional smoke transport and local urban building sources is a problem that does not yet have a cleanly integrated solution. A nested grid is required, from global (50 km) through regional (10 km), mesoscale (1 km), to urban CFD (10 m). Experimental stacks exist; production-class systems remain a frontier.

Where Ukraine stands — and what WildFiresUA does

Ukraine has no national atmospheric chemistry service with operational air-quality forecasting at the level of CAMS or NOAA. Existing infrastructure consists of separate monitoring stations at EcoCity and Sensor.Community, plus state monitoring through the Hydrometeorological Centre; the modelling component is absent at the national level. WildFiresUA fills this niche specifically for wildfire smoke: an operational WRF (1 km) + FLEXPART stack with daily updates, providing a 72-hour smoke transport forecast for the entire territory of Ukraine.

The architectural choice — pure Lagrangian WRF+FLEXPART — is dictated by the computational budget. Eulerian CMAQ or WRF-Chem on a 1 km grid requires 10-30× more HPC resources; we currently lack such infrastructure. Partnership with FMI (via SILAM for cross-validation) and with ULCO (via MesoNH-ForeFire for post-event analyses) is the strategic response to this gap.

CALPUFF is a separate stack for nuclear-plant radiological scenarios. It is the standard model for the IAEA-approved approach to assessing impact on protected areas; we use it primarily for post-event analyses and exercises with DSES Ukraine in the framework of response planning for potential incidents at Zaporizhzhia, Khmelnytskyi, and South Ukraine NPPs.

Conclusion

Atmospheric dispersion of wildfire smoke is a mature field of computational atmospheric physics with two large methodological schools (Lagrangian: HYSPLIT, FLEXPART, SILAM, NAME, CALPUFF; Eulerian: WRF-Chem, CMAQ, EMEP MSC-W) and several hybrid systems (SILAM with both modes). No model “wins” unconditionally — the choice is dictated by task type, computational budget, and partnership availability. Ukraine, through WildFiresUA, positions itself as a WRF+FLEXPART stack with backup partnerships at FMI (SILAM) and ULCO (MesoNH-ForeFire), and CALPUFF for radiological scenarios. This is a working compromise that delivers 80% of the operational result of a full CAMS-class national stack at 10-20% of the budget.

Ukrainian startup ecosystem: follow TechUkraine and AIN.ua — the two leading outlets covering Ukrainian deep tech, climate tech, and environmental startups.

What to do today

Check the YourAirTest air quality map for your city — recent PM2.5 readings.
If the topic resonates, share with colleagues; we monitor referrals via Google Search Console.
To contribute data (sensor measurements, regional models) reach out via the contact form.