The Hybrid Single-Particle Lagrangian Integrated Trajectory model (HYSPLIT), developed by NOAA’s Air Resources Laboratory, is one of the most widely used models for atmospheric trajectory and dispersion calculations. We present the model’s historical evolution over the last 30 years from simple hand-drawn back trajectories to very sophisticated computations of transport, mixing, chemical transformation, and deposition of pollutants and hazardous materials. We highlight recent applications of the HYSPLIT modeling system, including the simulation of atmospheric tracer release experiments, radionuclides, smoke originated from wild fires, volcanic ash, mercury, and wind-blown dust.
This work presents HYSPLIT’s historical evolution over the last three decades along with recent model developments and applications.
The National Oceanic and Atmospheric Admin-istration (NOAA) Air Resources Laboratory’s(ARL) Hybrid Single-Particle Lagrangian Integrated Trajectory model (HYSPLIT) (Draxler and Hess 1998) is a complete system for computing simple air parcel trajectories as well as complex transport, dispersion, chemical transformation, and deposition simulations. HYSPLIT continues to be one of the most extensively used atmospheric transport and dispersion models in the atmospheric sciences community [e.g., more than 800 citations to Draxler and Hess (1998) on Web of Science; http://thomsonreuters.com/thomson-reuters-web-of-science/]. One of the most common model applications is a back-trajectory analysis to determine the origin of air masses and establish source–receptor relationships [Fleming et al. (2012) and references therein]. HYSPLIT has also been used in a variety of simulations describing the atmospheric transport, dispersion, and deposition of pollutants and hazardous materials. Some examples of the applications (Table 1) include tracking and forecasting the release of radioactive material (e.g., Connan et al. 2013; Bowyer et al. 2013; H. Jeong et al. 2013a), wildfire smoke (e.g., Rolph et al. 2009), wind-blown dust (e.g., Escudero et al. 2011; Gaiero et al. 2013), pollutants from various stationary and mobile emission sources (e.g., Chen et al. 2013), allergens (e.g., Efstathiou et al. 2011), and volcanic ash (e.g., Stunder et al. 2007).
The model calculation method is a hybrid between the Lagrangian approach, using a moving frame of reference for the advection and diffusion calculations as the trajectories or air parcels move from their initial location, and the Eulerian methodology, which uses a fixed three-dimensional grid as a frame of reference to compute pollutant air concentrations (the model name, no longer meant as an acronym, originally reflected this hybrid computational approach). The HYSPLIT model has evolved throughout more than 30 years, from estimating simplified single trajectories based on radiosonde observations to a system accounting for multiple interacting pollutants transported, dispersed, and deposited over local to global scales. In this paper we walk the reader through the model’s history describing the ideas that inspired its inception, the evolution of the scientific concepts and parameterizations that were incorporated into successive model versions, and the most recent innovations.
MODEL HISTORICAL BACKGROUND: 1940s–70s.
The scientific foundation and inspiration for HYSPLIT’s trajectory capabilities can be traced back to 1949 (Fig. 1), when the Special Project Section (SPS) (ARL’s predecessor) of the U.S. Weather Bureau [now NOAA’s National Weather Service (NWS)] was charged with trying to find the source of radioactive debris originating from the first Soviet atomic test and detected by a reconnaissance aircraft near the Kamchatka Peninsula. For that purpose, back trajectories were calculated by hand based on wind data derived from twice-daily radiosonde balloon measurements. These trajectories followed 500-hPa heights assuming geostrophic wind flow. Although these back trajectories were calculated more than 60 years ago, the percentage error between the calculated and actual source location relative to the distance covered by the trajectories was remarkably low (about 5%; Machta 1992). Since then, trajectory calculations have been one of the backbones of ARL’s research activities (e.g., Angell et al. 1966, 1972, 1976).
During the mid-1960s, Pasquill (1961) and Gifford (1961) described the estimation of the horizontal and vertical standard deviation of a continuous plume concentration distribution, which constituted the basis for the construction of the so-called Gaussian dispersion models. One such model was developed at ARL (Slade 1966, 1968; Fig. 1) based on data from the well-known Project Prairie Grass (Barad 1958). Using this Gaussian approach and assuming steady state with homogeneous and stationary turbulence, air concentrations were estimated based on wind data collected at a single site. Extending this work in the late 1960s and early 1970s to handle more realistic (changing) weather conditions, ARL scientists developed the Mesoscale Diffusion (MESODIFF) model (Start and Wendell 1974; Fig. 1) in response to health and safety concerns at the Idaho National Reactor Testing Station (NRTS) related to planned or accidental releases of radioactive material into the atmosphere. This segmented Gaussian puff model used gridded data interpolated from a network of 21 tower-mounted wind sensors located within the boundaries of the NRTS (Wendell 1972) to account for spatial variability of the horizontal wind flow near the surface. The simulations used a maximum of 400 puffs and incorporated time-varying diffusion rates. A similar approach was used during the mid-1970s, when ARL researchers (Heffter et al. 1975) combined trajectories with a Gaussian plume model to compute long-range air concentrations from gaseous or particulate emissions from a uniform, continuous point source based on rawinsonde data.
HYSPLIT DEVELOPMENT HISTORY: 1980s–2000s.
These previous dispersion studies established the scientific basis for the development of HYSPLIT, version 1 (HYSPLIT1), in the early 1980s (Draxler and Taylor 1982; Fig. 1). In this initial version, segmented pollutant puffs were released near the surface and their trajectories were followed for several days. Transport was calculated from wind observations based on rawinsonde data (not interpolated) taken twice daily. Assumptions included no vertical mixing at night and complete mixing over the planetary boundary layer (PBL) during the day. Nocturnal wind shear was modeled by vertically splitting the puffs that extended throughout the PBL into 300-m subpuffs during the nighttime transport phase of the calculation. The model was used to simulate Kr-85 released to the atmosphere and sampled at multiple locations in the Midwestern United States during a 2-month field experiment (Draxler 1982). Later on, HYSPLIT version 2 (HYSPLIT2; Fig. 1) included the use of interpolated rawinsonde or any other available measured data to estimate vertical mixing coefficients that varied in space and time (Draxler and Stunder 1988). These mixing coefficients were derived from the Monin–Obukhov length, friction velocity, and surface friction potential temperature (Draxler 1987). This model version was applied to simulate the Cross Appalachian Tracer Experiment (CAPTEX; Ferber et al. 1986). Before the early 1990s, HYSPLIT only used rawinsonde observations with very limited spatial (e.g., 400 km) and temporal (e.g., 12 h) resolution for the calculation of transport and dispersion. Not until development of HYSPLIT version 3 (HYSPLIT3; Fig. 1) did the model utilize gridded output from meteorological models such as the Nested Grid Model (NGM; Table 2). HYSPLIT3 allowed the calculation of trajectories as well as transport and dispersion of pollutants using cylindrical puffs that grow in time and split when reaching the grid size of the meteorological data (Draxler 1992). This version was applied to simulate the Across North America Tracer Experiment (ANATEX; Draxler and Heffter 1989). Chemical formation and deposition of sulfate ( ; Rolph et al. 1992, 1993a) were incorporated into HYSPLIT3 in a first attempt to include chemistry in the modeling system. This application incorporated gas- and aqueous-phase oxidation of sulfur dioxide (SO2) and dry and wet removal of SO2 and aerosol . In HYSPLIT3, chemical transformations occurred only within each Lagrangian puff, without any interaction with other puffs. As part of a broader acid precipitation research effort, the model was applied over the eastern United States for 1989 and compared against observed seasonal and annual spatial patterns of SO2 and concentrations in air and wet deposition of from precipitation. HYSPLIT3 was also used to model the atmospheric fate and transport of semivolatile (SV) pollutants by incorporating a dynamic vapor/particle partitioning algorithm and including chemical transformations initiated by the hydroxyl radical (OH) and photolysis. HYSPLIT-SV has been used to estimate the transport and deposition of polychlorinated dibenzo-p-dioxin and polychlorinated dibenzofurans (PCDD/F) to the Great Lakes, including the estimation of detailed source–receptor relationships (Cohen et al. 1995, 1997b, 2002). Also, HYSPLIT-SV was utilized to obtain source–receptor results for atrazine transport and deposition to the Great Lakes and other sensitive ecosystems (Cohen et al. 1997a).
By the end of the 1990s, many new features had been incorporated into HYSPLIT version 4 (HYSPLIT4; Draxler and Hess 1998; Fig. 1), the basis for current model versions. The innovations include an automated method of sequentially using multiple meteorological grids going from finer to coarser horizontal resolution (e.g., Table 2) and the calculation of the dispersion rate from the vertical diffusivity profile, wind shear, and horizontal deformation of the wind field. HYSPLIT4 allows the use of different kinds of Lagrangian representations of the transported air masses: three-dimensional (3D) particles, puffs, or a hybrid of both. A 3D “particle” is a point, computational mass—representing a gaseous or particulate-phase pollutant—moved by the wind field with a mean and a random component (see section on “Transport, dispersion, and deposition calculation” and the online supplement, which can be found online at http://dx.doi.org/10.1175/BAMS-D-14-00110.2). Individual 3D particles never grow or split, but a sufficient number needs to be released to represent the downwind horizontal and vertical pollutant distribution. A single puff, on the other hand, represents the distribution of a large number of 3D particles by assuming a predefined concentration distribution (Gaussian or top hat) in the vertical and horizontal directions. They grow horizontally and vertically according to the dispersion algorithms for puffs [see Draxler and Hess (1998) for details], equivalent to the evolving distribution of particles in a comparable 3D particle calculation. Furthermore, puffs split if they become too large to be represented by a single meteorological data point. To avoid the puff number quickly reaching the computational array limits, puffs of the same age occupying the same location may be merged (see online supplement for details about splitting and merging). An alternative approach is to simulate the dispersion with a 2D object (planar mass, having zero vertical depth), where the contaminant has a puff distribution in the horizontal direction, growing according to the dispersion rules for puffs and splitting if it gets too large. In the vertical, however, they are treated as Lagrangian particles. This option permits the model to represent the more complex vertical structure of the atmosphere with the higher fidelity possible when using particles while representing the more uniform horizontal structure using puffs.
Version 4 of HYSPLIT has been the basis for the construction of essentially all model applications for the last 15 years. It has been evaluated against ANATEX observations and has been applied to estimate radiological deposition from the Chernobyl accident (Kinser 2001) and to simulate the Rabaul volcanic eruption (Draxler and Hess 1997, 1998). At the beginning of the 2000s, applications started to incorporate nonlinear chemical transformation modules to simulate ozone (O3) in the lower troposphere. Specific examples are given below, but in general, incorporating nonlinear chemical processes into a Lagrangian framework such as HYSPLIT constitutes a challenging task because there is no simple approach to deal with the chemical interactions that can occur among Lagrangian particles or puffs. The usual approach is to rely on a Lagrangian methodology to compute transport, dispersion, and deposition and an Eulerian framework to represent the chemical transformations of different reactive species (Chock and Winkler 1994b).
Draxler (2000) included a simplified photochemical scheme that describes the formation of tropospheric O3 using the integrated empirical rate (IER) chemical module (Johnson 1984; Azzi et al. 1995). The model configuration was very similar to that of Rolph et al. (1992), using a Lagrangian approach to simulate the transport, dispersion, and deposition and an Eulerian framework to calculate the O3 concentrations with no interaction among puffs. Using HYSPLIT driven by the Eta Data Assimilation System (EDAS) meteorological dataset (Table 2), this approach was applied to the area of Houston, Texas, for the summer of 1997.
A more generalized nonlinear chemistry module was incorporated into HYSPLIT (HYSPLIT CheM) to calculate the spatial and temporal distribution of different photochemical species in the lower troposphere over a regional scale (Stein et al. 2000). Using 3D particles, transport and dispersion were computed using meteorological fields from a mesoscale model [e.g., fifth-generation Pennsylvania State University–National Center for Atmospheric Research Mesoscale Model (MM5)]. Once the concentrations of all the chemicals were calculated over a regular Eulerian grid by summing the masses of the particles in the box where they reside, the Carbon Bond IV (Gery et al. 1989) mechanism was used to model chemical transformations. The resulting concentrations from the chemical evolution of each species were then redistributed as a change in the mass of each particle within the cell. It was assumed that the ratio of the final to initial concentration is equal to the ratio of the final to the initial mass for the corresponding species (Chock and Winkler 1994b; Stein et al. 2000). HYSPLIT CheM was applied to simulate O3 concentrations in Pennsylvania for a case study in 1996 (Stein et al. 2000). In addition, this particular model application constituted the first routine implementation of the particle-in-grid approach applied to the forecast of air quality (Kang et al. 2005).
At about the same time, HYSPLIT-SV was updated to the HYSPLIT4 framework and was used to estimate the transport and deposition of PCDD/F to the Great Lakes (Cohen et al. 2002) and to later estimate the fate and transport of PCDD/F emitted from the in situ burning of sea surface oil following the Deepwater Horizon spill (Schaum et al. 2010). In addition, HYSPLIT-SV was extended to simulate atmospheric mercury (HYSPLIT-Hg) by adding new gas-phase reactions and a treatment of aqueous-phase equilibrium and chemistry. Spatiotemporally resolved reactant concentrations (e.g., O3, OH, SO2) were estimated for each puff based on external model results and/or algorithms based on empirical data. This puff-based version of HYSPLIT-Hg has been used to analyze the transport and deposition of mercury to the Great Lakes from U.S. and Canadian sources (Cohen et al. 2004) and the fate and transport of mercury emissions in Europe (Ryaboshapko et al. 2007a,b).
By the end of the 2000s, new emissions features were included that are more sophisticated than the explicit input of an emission rate—that is, how much mass is assigned to each particle. Two predefined emission algorithms result in time-varying emissions based upon changing meteorological conditions. The first application for this approach, for wind-blown dust, relied on the calculated friction velocity and satellite-derived land use information to estimate the emissions. This algorithm (see online supplement) was applied to estimate dust levels over the contiguous United States (Draxler et al. 2010) and globally (Wang et al. 2011). The second time-varying emission algorithm allows for the estimation of plume rise using the buoyancy terms based on heat release, wind velocity, and friction velocity. This parameterization (online supplement) has been applied to predict transport and dispersion of smoke originating from forest fires over the contiguous United States (Rolph et al. 2009; Stein et al. 2009), Alaska, and Hawaii, and was used in the simulation of emissions from in situ sea surface oil burning (Schaum et al. 2010).
RECENT MODEL DEVELOPMENTS.
Transport, dispersion, and deposition calculation.
Many upgrades that reflect the most recent advances in the computation of dispersion and transport have been incorporated into HYSPLIT over the last 15 years. Only a brief introduction is given here; further details can be found in the online supplement. The computation of the new position at a time step (t + ∆t) due to the mean advection by the wind determines the trajectory that a particle or puff will follow. In other words, the change in the position vector Pmean with time
is computed from the average of the three-dimensional velocity vectors V at their initial and first-guess positions (Draxler and Hess 1998). Equation (1) is the basis for the calculation of trajectories in HYSPLIT. Only the advection component is considered when running trajectories. The turbulent dispersion component is only needed to describe the atmospheric transport and mixing processes for 3D particles and puffs.
The dispersion equations are formulated in terms of the turbulent velocity components. In the 3D particle implementation of the model, the dispersion process is represented by adding a turbulent component to the mean velocity obtained from the meteorological data (Fay et al. 1995); namely,
where Uʹ and Wʹ correspond to the turbulent velocity components, Xmean and Zmean are the mean components of particle positions, and Xfinal and Zfinal are the final positions in the horizontal and vertical, respectively. The turbulence component is always added after the advection computation.
Here, Uʹ and Wʹ are calculated based on the modified discrete-time Langevin equation [Chock and Winkler (1994a) and references therein], which is expressed as a function of the velocity variance, a statistical quantity derived from the meteorological data, and the Lagrangian time scale. Four updated model parameterizations are available for the calculation of vertical mixing—namely, i) assuming the vertical mixing diffusivities follow the coefficients for heat, ii) based on the horizontal and vertical friction velocities and the PBL height, iii) using the turbulent kinetic energy (TKE) fields, or iv) directly provided by the meteorological data. Furthermore, a very simplified enhanced mixing deep convection parameterization based on the convective available potential energy has been included (see the online supplement).
The description of how wet and dry deposition is simulated in HYSPLIT can be found elsewhere (Draxler and Hess 1998). However, a new option for in-cloud wet scavenging parameterization has been incorporated into the modeling system based on the estimation of a scavenging coefficient (Leadbetter et al. 2015; see the online supplement for further details).
Embedded global Eulerian model (GEM) and multiple Lagrangian representations.
Recently, a global Eulerian model (GEM; Draxler 2007) was included as a module of the HYSPLIT modeling system. In this option, particles or puffs are first released in the Lagrangian framework and carried within HYSPLIT until they exceed a certain age at which point their mass is transferred to the GEM. Currently, the GEM can only be driven by global meteorological data from the Global Forecast System (GFS), Global Data Assimilation System (GDAS), or the National Centers for Environmental Prediction (NCEP)–National Center for Atmospheric Research (NCAR) reanalysis models (Table 2). The GEM includes the following processes: advection, horizontal and vertical mixing, dry and wet deposition, and radioactive decay. One advantage of this approach is that near the emission sources the Lagrangian calculation better depicts the details of the plume structure, without the initial artificial diffusion of an Eulerian model. Then, when the plume features no longer need to be resolved, the particle or puff mass is transferred to the Eulerian framework. The GEM calculation methodology is primarily aimed at improving the computational efficiency of global transport applications and especially in situations when shorter-range plumes may interact with hemispheric or global background variations. In such cases, a single simulation can be used to efficiently model global transport, dispersion, and deposition that would be computationally burdensome using 3D particles or puffs alone considering the number of such particles or puffs that would be required for global coverage. In particular, this approach is very effective in simulating the initiation of individual dust storm plumes that quickly merge into a regional or even hemispheric event (e.g., Wang et al. 2011). In addition, the mercury analysis performed to estimate North American sources influencing the Great Lakes area has recently been extended using the GEM capability within HYSPLIT to include sources worldwide (Cohen et al. 2011, 2014).
Besides transferring the mass from puff/particles to the GEM, HYSPLIT recently incorporated an option to allow the Lagrangian description to change between particles and puffs during the transport process depending upon their age (since release); this is called the mixed-mode approach. This option attempts to avoid dealing with a large number of particles while keeping the best physical description of the phenomenon under study. For instance, a mixed mode may be selected to take advantage of the more accurate representation of the 3D particle approach near the source and the horizontal distribution information provided by one of the hybrid puff approaches at longer transport distances. The following transformation options are available in the HYSPLIT system: 3D particle converting to Gaussian horizontal puff and vertical particle distribution (Gh-Pv), 3D particle converting to top-hat horizontal puff and vertical particle distribution (THh-Pv), Gh-Pv converting to 3D particle, THh-Pv converting to 3D particle, and 3D particle or puffs (Gaussian or top hat) converting to the GEM. In general, fewer Lagrangian particles/puffs are needed under the mixed-mode approach for a given level of computational resolution.
Source estimations using footprints.
Back-trajectory calculations have been one of the most attractive and prominent features by which HYSPLIT has been used in many studies [Fleming et al. (2012) and references therein]. Although trajectories offer a simple assessment of source–receptor relationships, a single trajectory cannot adequately represent the turbulent mixing processes that air parcels experience during transport. However, coupling the back-trajectory calculation with a Lagrangian dispersion component can produce a more realistic depiction of the link between the concentrations at the receptor and the sources influencing it (Stohl et al. 2002; Lin et al. 2003). To this end, backward-in-time advection with dispersion has been included in the HYSPLIT modeling system by simply applying the dispersion equations to the upwind trajectory calculation [i.e., Eqs. (2) and (3) are assumed to be reversible when integrated from t + ∆t to t]. Under this approach, the increasingly wider distribution of Lagrangian particles or puffs released from a receptor undergoing backward-in-time transport and dispersion represents the geographical extent and strength of potential sources influencing the location of interest. Nevertheless, this particular model application must satisfy the well-mixed criteria, include appropriate representation of the interaction between the wind shear and vertical turbulence, and provide for sufficient decay in the autocorrelation of Uʹ [see Lin et al. (2003) for details]. In addition, the mean trajectory component of the calculation, which is normally considered to be reversible, should not intersect the ground; otherwise it loses information and becomes irreversible. From the practical computational perspective, performing backward calculations from a few receptor points is more efficient than forward calculations from many more potential source locations to find the best match with the receptor data even at the loss of some accuracy. The more accurate forward calculation can be used once a smaller set of source locations have been identified.
Examples of the use of HYSPLIT’s backward Lagrangian dispersion modeling methodology include the estimation of mercury sources impacting New York State (Han et al. 2005), backtracking anthropogenic radionuclides using a multimodel ensemble (Becker et al. 2007), and quantifying the origins of carbon monoxide and ozone over Hong Kong (Ding et al. 2013). In addition, this approach has been adopted in other applications that have used HYSPLIT as the foundational model for back dispersion calculations [e.g., the Stochastic Time-Inverted Lagrangian Transport (STILT) model: Lin et al. 2003; Wen et al. 2012]. Building upon this capability, HYSPLIT also allows for the direct calculation of source footprints (Lin et al. 2003), defined as areas of surface emission fluxes that contribute to changes in concentrations at a receptor (online supplement). Examples of the use of this approach can be found elsewhere (e.g., Gerbig et al. 2003; Kort et al. 2008; Zhao et al. 2009; S. Jeong et al. 2013).
Pre- and postprocessors.
The preparation of the required meteorological input data and the analysis of the simulation outputs are important additional aspects to consider when running the HYSPLIT model. HYSPLIT includes a series of preprocessors designed to prepare model-ready meteorological gridded datasets and postprocessors to analyze multiple trajectory outputs and concentration ensembles.
Meteorological model preprocessing
HYSPLIT can use a large variety of meteorological model data in its calculations, ranging from mesoscale to global scales. Rather than having a different version of HYSPLIT to cope with the variations in variables and structure for each meteorological data source, a customized preprocessor is used to convert each meteorological data source into a HYSPLIT-compatible format. In this way HYSPLIT can easily be run with one or more meteorological datasets at the same time, using the optimal data for each calculation point. Table 2 describes some of the meteorological data already formatted for HYSPLIT that are publicly available from NOAA ARL (www.ready.noaa.gov/archives.php) and NOAA NCEP (ftp://ftpprd.ncep.noaa.gov//pub/data/nccf/com/hysplit/prod/). In addition, the following model outputs can also be used to drive HYSPLIT: the Weather Research and Forecasting (WRF) Model (Skamarock et al. 2008), MM5 (Grell et al. 1994), the Regional Atmospheric Modeling System (RAMS; Pielke et al. 1992), and the European Centre for Medium-Range Weather Forecasts (ECMWF) interim reanalysis (ERA-Interim; Dee et al. 2011).
Multiple trajectory analysis
The calculation of forward and backward trajectories allows for the depiction of airflow patterns to interpret the transport of pollutants over different spatial and temporal ranges. Frequently trajectories are used to track the airmass history or to forecast airmass movement and to account for the uncertainty in the associated wind patterns. Grouping trajectories that share some commonalities in space and time simplifies their analysis and interpretation and also reduces the uncertainty in the determination of the atmospheric transport pathways (Fleming et al. 2012).
Once the multiple trajectories representing the flow pattern of interest have been calculated, trajectories that are near each other can be merged into groups, called clusters, and represented by their mean trajectory. Differences between trajectories within a cluster are minimized while differences between clusters are maximized. Computationally, trajectories are combined until the total variance of the individual trajectories about their cluster-mean starts to increase substantially (Stunder 1996). This occurs when disparate clusters are combined. For references to cluster analysis methods the reader is referred to, for example, Borge et al. (2007), Karaca and Camci (2010), Markou and Kassomenos (2010), Baker (2010), and Cabello et al. (2008).
The use of dispersion model ensembles—with the objective of improving plume simulations and assessing their uncertainty—has been an increasingly attractive approach to study atmospheric transport (e.g., Potempski et al. 2008; Lee et al. 2009; Solazzo et al. 2013; Stein et al. 2015). The HYSPLIT system has a built-in capability to produce three different simulation ensembles. This ensemble approach has been applied to case studies using different sets of initial conditions and internal model physical parameters (Draxler 2003; Stein et al. 2007; Chen et al. 2012). These built-in ensembles are not meant to be comprehensive and only account for some of the components of the concentration uncertainty, such as those arising from differences in initial conditions and model parameterizations. The first, called the “meteorological grid” ensemble, is created by slightly offsetting the meteorological data to test the sensitivity of the advection calculation to the gradients in the meteorological data fields. The rationale for the shifting is to assess the effect that a limited spatial- and temporal-resolution meteorological data field—an approximation of the true flow field which is continuous in space and time—has on the output concentration (Draxler 2003). The second, called the “turbulence” ensemble, represents the uncertainty in the concentration calculation arising from the model’s characterization of the random motions created by atmospheric turbulence (Stein et al. 2007). This ensemble is generated by varying the initial seed of the random number generator used to simulate the dispersive component of the motion of each particle. The model already estimates this turbulence when computing particle dispersal. However, normally, a sufficiently large number of particles would be released to ensure that each simulation gives similar results. In the turbulence ensemble approach, the number of particles released is reduced and multiple simulations are run, each with a different random number seed. The third, the “physics” ensemble, is built by varying key physical model parameters and model options such as the Lagrangian representation of the particles/puffs, Lagrangian time scales, and vertical and horizontal dispersion parameterizations.
MODEL EVALUATION USING TRACER EXPERIMENTS.
Atmospheric tracer experiments offer a unique opportunity to evaluate the transport and dispersion independently from other model components such as chemical transformations or deposition. In these experiments, known amounts of an inert gas are emitted into the atmosphere and measured downwind for several days. One such experiment is the CAPTEX (Ferber et al. 1986) campaign that took place from 18 September to 29 October 1983 and consisted of six 3-h perfluro-monomethylcyclohexane (PMCH) releases: four from Dayton, Ohio, and two from Sudbury, Ontario, Canada. Samples were collected at 84 sites located 300–800 km downwind from the source at 3- and 6-h averaging periods for approximately 2–3 days after each release.
To illustrate how HYSPLIT’s updates are evaluated, we performed a series of simulations using some of the updated vertical mixing parameterizations and compared the results with CAPTEX data. HYSPLIT was configured to simulate each of the six CAPTEX experiments by releasing 50 000 3D Lagrangian particles, using a vertical Lagrangian time scale (TLw)—which is a measure of the persistence of fluid motion—of 5 and 200 s for stable and unstable conditions, respectively (see online supplement for further details), and employing an output concentration grid over the relevant domain with dimensions 0.25° × 0.25° by 100 m in depth. The fluxes of heat and momentum from the meteorological model were used to estimate the boundary layer stability parameters. All the HYSPLIT simulations used the meteorological data fields from the Advanced Research version of WRF (Skamarock et al. 2008), version 3.5. The innermost WRF domain was configured to cover the northeastern U.S. with a horizontal resolution of 9 km and 27 vertical layers using the Mellor–Yamada–Janjić (MYJ) (Janjić 1990) PBL scheme. The MYJ parameterization is a local, 1.5-order closure in which the TKE is a prognostic variable that is used for determining the diffusion coefficients. Moreover, as an alternative to the instantaneous wind fields generally used to drive the transport and dispersion simulations, WRF also produces time-averaged, mass-coupled, and horizontal and vertical velocities (Nehrkorn et al. 2010; Hegarty et al. 2013) that are automatically used by HYSPLIT if they are available.
Because of the difficulty in determining model performance using a single evaluation metric, we evaluate the model’s performance against observations using the ranking method as defined by Draxler (2006). This method adds the correlation coefficient R, fractional bias (FB), figure of merit in space (FMS), and Kolmogorov–Smirnov parameter (KSP) into a single normalized Rank parameter that ranges from 0 to 4 (from worst to best); namely,
Table 3 shows the model performance for six CAPTEX releases using four different model configurations that combine two available mixing calculation options: one based on the horizontal and vertical friction velocities and the PBL height (Kantha and Clayson 2000) and another using TKE from WRF. These are combined with the use of snapshot or time-averaged mass-coupled wind fields. Note that no particular combination gives the best model performance for all the tracer releases, indicating that different parameterizations present an advantage or disadvantage under different atmospheric conditions. For example, Figs. 2a and 2b compare the simulated and observed tracer concentrations from CAPTEX tracer releases 2 and 7, showing that the model captures the characteristics of the geographical distribution and magnitude of measured PMCH concentrations. The reader is referred to Hegarty et al. (2013) for a more detailed and complete comparison with CAPTEX and additional tracer data. Other transport and dispersion models such as STILT (Lin et al. 2003) and Flexible Particle dispersion model (FLEXPART; Stohl et al. 2005) were compared in that work and showed similar performance.
We strongly believe that comparing with tracer experiments should be an integral part of the evaluation of transport and dispersion models. To facilitate this, we have made the CAPTEX data—along with an additional nine other tracer datasets (most consisting of multiple releases)—publically available from the Data Archive of Tracer Experiments and Meteorology (DATEM; www.arl.noaa.gov/DATEM.php) in a common format. DATEM also contains HYSPLIT simulation results for each experiment.
The HYSPLIT modeling system can be currently run on PC, Mac, or Linux platforms using a single processor. Multiple processor parallelized environment calculations based on a message passing interface (MPI) implementation are available for Mac and Linux. The system includes a suite of pre- and postprocessing programs to create input data as well as to visualize and analyze the simulation outputs. These programs can be called through a graphical user interface (GUI), the command line, or automated through scripts. The model is available for download at www.ready.noaa.gov/HYSPLIT.php. A registered version is also available that adds the capability of running the model with current (today’s) forecast meteorological data. About 3,000 registered users have already downloaded HYSPLIT. The source code is available upon request following the instructions to download the registered version.
Another way to gain public access to meteorological data and run HYSPLIT trajectory and dispersion simulations is through the Real-Time Environmental Applications and Display System (READY) (Rolph et al. 1993b), a web-based system developed and maintained by ARL (ready.arl.noaa.gov/). READY brings together the trajectory and dispersion model, graphical display programs, and textual forecast programs generated over many years at ARL into a particularly easy-to-use form. Since its initial development in 1997 (Fig. 1), thousands of users (about 80,000 HYSPLIT simulations per month) have generated products from READY for their day-to-day needs and research projects. In addition, a specialized website has been developed to allow NWS forecasters to run HYSPLIT for local events (e.g., hazardous materials incidents, forest fires, and nuclear accidents) and relay the results directly to state and local emergency managers through a customized web page.
Every year HYSPLIT developers offer training workshops on the installation and use of the modeling system, including a wide variety of applications such as volcanic eruptions, radionuclide accidents, dust storms, wildfire smoke, and tracer experiments. Training materials, including a self-paced tutorial, are available at www.arl.noaa.gov/HYSPLIT_workshop.php. Workshop participants typically include members of the U.S. and international governments, private industry, and academia. In addition, a forum for HYSPLIT model users is available to communicate questions, problems, and experiences (https://hysplitbbs.arl.noaa.gov/). This forum currently has more than 2,000 participants.
Many of the model applications described in this work are currently being used to fulfill ARL’s mission. One of the many functions of ARL is to provide atmospheric transport and dispersion information and related research to NOAA, other federal agencies, and the general public in order to estimate the consequences of atmospheric releases of pollutants, radioactivity, and other potentially harmful materials.
For example, ARL’s volcanic ash model [initially Volcanic Ash Forecast Transport and Dispersion (VAFTAD; Heffter and Stunder 1993); now HYSPLIT (Stunder et al. 2007)] (Fig. 3) provides critical information on plume transport and dispersion to the aviation industry (www.ready.noaa.gov/READYVolcAsh.php). HYSPLIT is currently run operationally by the NOAA/NWS to forecast the transport and dispersion of volcanic ash in and near the U.S. Volcanic Ash Advisory Centers’ (VAAC) areas of responsibility covering North and Central America. Meteorologists at the VAAC and the Meteorological Watch Offices use the HYSPLIT forecasts, among other sources of information, for writing Volcanic Ash Advisories and Significant Meteorological Information warning messages (called SIGMETs). The HYSPLIT dispersion forecasts are issued to the public and made available online, such as at the NWS Aviation Weather Center (http://aviationweather.gov/iffdp/volc). Additional volcanic ash applications of the model include HYSPLIT’s participation in a dispersion model intercomparison among the international centers that provide advisories for aviation (Witham et al. 2007), investigating source-term sensitivity (Webley et al. 2009), locating the volcano source given downwind ash–aircraft encounters (Tupper et al. 2006), and modeling VOG (a mixture of SO2 and sulfate) in Hawaii (http://mkwc.ifa.hawaii.edu/vmap/index.cgi).
As a result of communications difficulties between countries following the Chernobyl accident in the spring of 1986, the World Meteorological Organization (WMO) was requested by the International Atomic Energy Agency (IAEA) and other international organizations to arrange for early warning messages about nuclear accidents to be transmitted over the Global Telecommunications System. In addition, some WMO member countries lacking extensive forecasting capability requested specialized pollutant transport and dispersion forecasts during these emergencies. Consequently, Regional Specialized Meteorological Centers (RSMCs; www.wmo.int/pages/prog/www/DPFSERA/EmergencyResp.html) were set up to respond to these needs. ARL, together with NOAA’s NCEP, constitute the Washington RSMC for transport and dispersion products through WMO. RSMC Washington, along with RSMC Montreal (operated by the Canadian Meteorological Centre), provide meteorological guidance and dispersion predictions using their respective models in the event of an atmospheric release of radioactive or hazardous materials crossing international boundaries in North, Central, and South America (www.arl.noaa.gov/rsmc.php). Furthermore, HYSPLIT was used to evaluate the consequences of the accidental release of nuclear material into the atmosphere from the Fukushima Daiichi Nuclear Power Plant following an earthquake and tsunami in March 2011 (e.g., Fig. 4; Draxler and Rolph 2012; Draxler et al. 2013).
Transport of forest fire smoke and its effect on weather has been a topic of NOAA interest at least since the middle of the last century (Smith 1950) and modeling the movement of smoke from large wildfires has been an ongoing development activity of ARL since 1998 (Rolph et al. 2009). This research eventually led to the first operational smoke forecasts over the continental U.S. in 2007 by NOAA in support of the National Air Quality Forecast Capability (Rolph et al. 2009) (www.arl.noaa.gov/smoke.php). Today, in addition to the continental United States, smoke forecasts are produced for Alaska and Hawaii on a daily basis to provide guidance to air quality forecasters and the public on the levels of particulate matter with diameters smaller than 2.5 µm (PM2.5) in the air (http://airquality.weather.gov/).
Finally, HYSPLIT has very recently been coupled inline to WRF (Ngan et al. 2015) taking advantage of the higher temporal frequency available from the meteorological data. The model runs within the WRF architecture using the same spatial and temporal resolution and it has been tested against CAPTEX and other tracer experiments. This is a very promising approach for applications influenced by rapidly changing conditions and/or complex terrain. Further evaluation of this approach is underway.
The authors thank Hyun-Cheol Kim for help with graphics and Dian Seidel for very helpful editorial comments. We also thank the many model users that have contributed to the development and improvement of HYSPLIT over the past three decades.
Publisher’s Note: This article was modified on 5 February 2016 to correct the empirical formula for sulfate.
A supplement to this article is available online (10.1175/BAMS-D-14-00110.2)