Influence of Dust on Precipitation During Landfalling Atmospheric Rivers in an Idealized Framework

Abstract

Atmospheric rivers (ARs) can provide as much as 50% of the total annual rainfall to the United States West Coast via orographic precipitation. Dust is thought to enhance orographic precipitation via the “seeder-feeder” mechanism, in which ice particles from a high cloud fall through a lower orographic cloud, seeding precipitation in the low cloud. Using the Weather Research and Forecasting model, we vary dust concentrations in simulations of two-dimensional flow over a mountain. This idealized framework allows us to test the sensitivity of the precipitation-dust response to a variety of different dust concentrations and initial conditions. The model is run using an ensemble of 60 radiosondes collected from Bodega Bay, CA in 2017–2018, clustered based on their vertical moisture profile into “deep moist,” “shallow moist,” and “subsaturated” clusters. The principle impact on precipitation is to increase the ratio of precipitation falling as snow. This produces a “spillover” effect, decreasing precipitation upwind of the peak and increasing precipitation downwind of the peak. The largest impacts on the snow/rain ratio occur at the end of the event, during cold front passage. The ensemble mean does not produce a significant seeder-feeder response, however, in individual cases with favorable initial conditions there is a significant increase in precipitation throughout the domain due to dust effects on the seeder-feeder mechanism. These findings afford an opportunity to build a more comprehensive understanding for the conditions under which dust aerosol can have a significant impact on precipitation during ARs, with implications for future developments in forecasting.

Key Points

• Dust increases the percentage of precipitation falling as snow/graupel during landfalling atmospheric rivers

• Increases in dust tend to decrease orographic precipitation upwind of the peak and increase orographic precipitation downwind of the peak

• The sensitivity of precipitation to dust depends on the initial thermodynamic structure of the atmosphere

1 Introduction

The United States West Coast can get as much as 50% of its total annual precipitation from a few large storm systems, known as atmospheric rivers (ARs) (Dettinger et al., 2011). ARs are characterized by long narrow bands of moisture where the vertically integrated water vapor transport (IVT) from the surface to 300 hPa is ≥250 kg m⁻¹ s^‒1 (Ralph et al., 2004; Rutz et al., 2014; Zhu & Newell, 1998). ARs are generally associated with a parent extratropical cyclone, with the AR core (region of maximum IVT) roughly aligned with the cold front of the extratropical cyclone. As the AR makes landfall, the typical progression is the passage of the warm front, followed by the AR core which is associated with the most intense precipitation, and then the passage of the cold front. Although IVT values generally drop off after the passage of the cold front, there can still be periods of intense precipitation after the cold front passes. A landfalling AR can produce intense precipitation lasting anywhere from hours to days (Dettinger et al., 2011). The bulk of this precipitation occurs due to orographic processes as the moist air mass of the AR is lifted, first by the coastal range, and then by the Sierras in California.

Most ARs are beneficial for the United States West Coast, increasing the availability of water storage and snowpack, but the most extreme events can lead to hazards such as floods and debris flow (Dettinger et al., 2011; Oakley et al., 2017; Ralph et al., 2006, 2019). As such, accurately forecasting the precipitation amount, intensity, and type is critically important for water managers in the region. The CalWater campaign (Cordeira et al., 2017; Ralph et al., 2016) was a multiyear series of field experiments between 2009 and 2018 targeted toward improving our scientific understanding and ability to forecast landfalling ARs. Using a combination of targeted research flights, ship and ground based measurements, the CalWater campaigns provided a wealth of data on the structure and intensity of ARs, as well as providing information on the distribution and type of aerosols, including dust and marine aerosols within the AR (Ault et al., 2011; Creamean et al., 2013).

Dust can influence orographic precipitation via its effect on ice nucleation processes (Ault et al., 2011; Creamean et al., 2013; Vali et al., 2015). In mixed phase clouds, such as those seen in ARs, ice primarily forms via heterogenous nucleation in which cloud water and/or water vapor condenses, deposits, and/or freezes onto an ice nucleus. Dust is one of the most abundant and effective types of ice nuclei (Atkinson et al., 2013; DeMott et al., 2003; Hande et al., 2015; Heintzenberg et al., 1996; Muhlbauer & Lohmann, 2009). Cornwell et al. (2019) analyzed in situ measurements of ice nucleating particles (INPs) at coastal sites in California and found that while sea spray aerosols were more abundant in the ambient air, mineral dust particles were the most abundant in ice crystal residuals, that is, that far more ice crystals nucleated around dust particles than sea spray aerosols. Ault et al. (2011) compared two ARs that made landfall in California in the winter of 2009. The storms had similar characteristics in terms of orientation and IVT maximum, but the second storm produced 1.4 times the precipitation of the first storm. Measurements collected during the CalWater Early Start observational campaign (Ralph et al., 2016) showed that the second storm contained a high concentration of long range transported dust. The authors found that the droplet size was significantly larger in the second storm, and hypothesized that the enhanced precipitation in the second storm was driven by increased dust concentrations. Subsequently, in the CalWater-1 field campaign (Ralph et al., 2016), Creamean et al. (2013) found evidence of dust influencing the “seeder-feeder mechanism,” in which ice forms in a mid-level “seeder” cloud, and then falls into and becomes rimed in a lower level “feeder” cloud. These hydrometeors then either precipitate as snow/graupel, or melt into liquid droplets. In the presence of supercooled water droplets, ice crystals will preferentially grow more quickly than liquid water droplets. Therefore, it is expected that the seeder-feeder mechanism will produce larger droplet sizes and more intense precipitation. Creamean et al. (2015) found that dust and biological particles both served as INPs in storms that made landfall over the northern Sierras in the winters of 2009, 2010, and 2011. Dust and biological INPs were typically found in storms with deep convective cloud systems, and biological INPs were most prominent in warm ARs. Creamean et al. (2016) found a similar relationship in the southern Sierras in the winters of 2011 and 2012. In a study of INPs found in precipitation samples during an AR in March 5–6, 2016, Martin et al. (2019) found a mixture of biological particles, dust, organic carbon, and marine aerosols acting as INPs. Samples were collected at a coastal site (Bodega Bay, CA, USA) and a site in the coastal mountain range (Cazadero, CA, USA). During this AR, the most abundant INPs were biological particles, with dust as the second most abundant. INP concentrations in the precipitation samples were enhanced in the early stages of the AR and following the passage of the cold front. Levin et al. (2019) demonstrated that in some storms, marine INPs can dominate, allowing ice to form at much warmer temperatures. Further research is needed to develop a comprehensive picture of the climatology of which aerosols are most important for ice formation processes during ARs.

A number of studies have demonstrated that pollution aerosols acting as cloud condensation nuclei (CCN) can affect orographic precipitation (e.g., Choudhury et al., 2019; Lynn et al., 2007; Muhlbauer et al., 2010; Saleeby et al., 2009). Relatively fewer studies have modeled the effects of dust on precipitation during specific storms. Hazra et al. (2016) used the MM5 model to study the effects of dust and black carbon aerosols on storms in the Himalayas and found that dust and soot aerosols lead to enhanced riming and increased precipitation efficiency. Using 3D simulations with the COSMOS model, Muhlbauer and Lohmann (2009) investigated the role of dust and black carbon in orographic precipitation and found that dust in particular led to an increased freezing rate in the orographic cloud. Precipitation in their model increased in the upslope and decreased on the downslope due to the earlier initialization of the ice phase. Fan et al. (2014) considered the role of dust and other aerosols during two case studies, February 16 and March 2, 2011. Using the WRF model over northern California, they found that dust significantly increased precipitation by as much as 15% over the Sierras during the February 16th AR, but had a much smaller impact on the March 2 event. Notably, the February 16 storm had a deep cloud layer, which formed after a shallow cloud merged with an elevated cloud layer on February 15. The cloud top temperature on the 16th was ‒36°C. In contrast, the March 2 event had a shallower cloud layer, with a cloud top temperature of only ‒20°C. Comparison with surface maps from the Weather Prediction Center (WPC) show that the cold front passed over northern California around 00Z on February 16 (WPC, 2019), 12 h before the start of the simulations, explaining the cooler cloud temperatures for this case. Fan et al. (2017) expanded on this analysis by considering a range of dust concentrations for the two cases. The authors found that in both cases, INPs tended to produce more precipitation, but in the cold case (February 16), it was due to enhanced deposition, whereas in the warm case (March 2nd) it was due to enhancements in both riming and deposition.

The ability to accurately forecast precipitation amount, intensity, location, and timing is critically important. In this work we perform a preliminary analysis of the sensitivity of orographic precipitation to dust under a range of different initial conditions using idealized two-dimensional WRF simulations, with the ultimate goal of incorporating dust into operational forecasts. The remainder of this article is organized as follows: Section 2 describes the model and data used. Section 3 presents the modeled precipitation response to changes in dust concentration, and discussion and conclusions are presented in Sections 5 and 6, respectively.

2 Data and Methods

2.1 Observations and Reanalysis

We test the sensitivity of orographic precipitation during landfalling ARs to increased dust concentration using the WRF model in an idealized two-dimensional setup (described in Section 2.3). The simulations are initialized at its western boundary using a subset (60) of the 245 radiosondes collected at Bodega Bay, CA, USA (star in Figure 1a), home to one of NOAA's AR Observatories, during the 2017–2018 Forecast Informed Reservoir Operations (FIRO) field campaign (Table 1; Jasperse et al., 2017; Ralph et al., 2021). Bodega Bay is situated at the mouth of the Russian River watershed, which is fed by the Lake Mendocino Reservoir, and gets 30%–50% of its annual rainfall from landfalling ARs (Dettinger et al., 2011; Ralph et al., 2013). Radiosondes are collected between mid-January and early April each year. The radiosondes collect data on temperature, relative humidity, and pressure as well as Global Positioning System data which is used to calculate wind speed and direction. During landfalling AR events, sondes are launched at 3 h intervals, going up to 1.5 h intervals during peak IVT conditions. The sondes typically collect data from near the surface (below 20 m) through the stratosphere. Sondes launched at 3 h intervals typically penetrate well into the stratosphere (upwards of 21 km) before the balloon pops, while sondes launched at 1.5 h intervals typically retrieve data up to the lower stratosphere (15 km) before being terminated. The high temporal density of observations allows us to evaluate the effects of dust on precipitation during different stages of an AR. The subset of 60 sondes was chosen to provide a large enough sample size to detect a signal out of the statistical noise, while still being a small enough sample to allow us to run a number of different scenarios without becoming too computationally expensive.

Figure 1

(a) Surface elevation of the Western U.S. The red star signifies the location of the Atmospheric River Observatory (ARO) in Bodega Bay, CA. The solid black line is a sample transect of a typical AR path. (b) Elevation along the transect (black) compared with the idealized model topography (blue), plotted as distance from the model’s western boundary.

Table 1. FIRO Radiosondes Collected at Bodega Bay, CA During Water Years 2017 and 2018 (Total Radiosondes Equals 245)

Month	No. of radiosondes
January 2017	57
February 2017	87
March 2017	16
April 2017	0
January 2018	29
February 2018	0
March 2018	37
April 2018	20

As an example, Figure 2 shows three sondes collected during the early, middle and late stages of the January 8–9, 2017 AR. This storm was a strong (AR4) event (Ralph et al., 2019). The first sonde (Figure 2a) was launched at 00Z on January 8, 2017. At this time in the storm the IVT over Bodega Bay was 384.0 kg m⁻¹ s⁻¹. The sonde is saturated in the lower troposphere, up to 850 hPa. There is a pronounced dry layer in the mid-troposphere. Above 400 hPa, the sonde remains subsaturated, but with a greater relative humidity, suggesting the possibility of forming ice. The winds at the surface are weak and predominantly southerly, strengthening, and transitioning to westerlies aloft. The second sonde (Figure 2b) was launched later the same day at 19:30Z. At this point, the AR core (the region of maximum IVT) was passing over Bodega Bay. The storm has a deep moist layer stretching into the mid-troposphere (500 hPa) and a calculated IVT of 1,086.9 kg m⁻¹ s⁻¹. The wind directions are consistent with Figure 2a, but the wind speeds have increased, particularly in the lower and mid troposphere. The third sonde (Figure 2c) was launched at 06Z on January 9th, after the cold front passed Bodega Bay (WPC, 2019). The IVT in this sonde dropped to 372.2 kg m⁻¹ s⁻¹. The atmosphere is saturated or near saturation up to 650 hPa, after which the sonde dries off dramatically. Unlike the earlier sondes, this sonde remains completely dry above 600 hPa. The surface winds have shifted to westerly flow and decreased in speed, as expected after the passage of a cold front. As we will show in Section 2.2, this structure is fairly typical of a landfalling AR.

Figure 2

Skew-T log-p for three radiosondes launched from Bodega Bay, CA during the January 8–9, 2017 atmospheric river event. The first sonde (a) was collected early in the event (January 8, 00Z). The second sonde (b) was collected near the peak observed integrated vapor transport (IVT) conditions at Bodega Bay (January 8, 19:30Z). The third sonde (c) was collected shortly after the cold front passed Bodega Bay (January 9, 06Z), as seen in comparisons with surface maps from the Weather Prediction Center (not shown). IVT is 384.0 kg m⁻¹ s⁻¹ initially (a), rises to 1086.9 kg m⁻¹ s⁻¹ (b), and then decreases back to 372.2 kg m⁻¹ s⁻¹ (c). The thick black lines are the in-situ temperature and the dashed black lines are the in-situ dew point temperature. All other lines and symbols assume their typical definitions.

In order to get a broader spatial picture of the development and positioning of the landfalling ARs considered here, we also utilize total column precipitable water from the ERA5 reanalysis dataset (Copernicus Climate Change Service (C3S), 2017) over the same time period covered by the radiosondes. ERA5 data is hourly on a 30 km grid with 137 vertical levels from the surface to 80 km. We refer to surface maps provided by the National Weather Service WPC for synoptic analysis (WPC, 2019).

2.2 Radiosonde Clusters

As discussed previously, Fan et al. (2014) examined the effect of dust on orographic precipitation and found evidence that the thermodynamic structure of the AR impacts the sensitivity of precipitation to dust. In order to further examine the potential role of the vertical structure of the AR on dust sensitivity, we classify the 245 radiosondes collected at Bodega Bay during the 2017–2018 FIRO field campaigns according to their vertical relative humidity profile using a k-means clustering algorithm. We interpolate the sondes to a common vertical grid with 50 m resolution. For our purposes, we are primarily interested in the moisture profile in the troposphere, so we restrict the clustering algorithm to relative humidity from 50 to 12,500 m. The lowest level of the interpolated sondes (0–50 m) is discarded due to missing data. The algorithm minimizes the euclidean distances between points in the same cluster, and calculates a centroid for each cluster. 20 sondes were removed from the analysis due to missing data. Of the remaining 225 sondes, we find three distinct clusters, shown in Figure 3. We use silhouette analysis (not shown) to determine that the choice of three clusters provides the most robust separation between clusters. Cluster one (blue) consists of 110 “deep moist” sondes. Radiosondes in this group are saturated or near saturated through the mid troposphere (up to 6,000 m). Figure 2b is an example of a deep moist sonde. Sondes in the second cluster (76, black) are saturated or near saturated in the lower troposphere (up to 3,000 m), and dry aloft (as in Figure 2c). The third and final cluster (red) is made up of 39 sondes that are subsaturated throughout the troposphere (Figure 2a is an example). However, this cluster was also the most variable, suggesting that to some extent it may represent sondes that do not cleanly fit into the first two clusters. The clusters will be referred to as “deep moist,” “shallow moist,” and “subsaturated” throughout the text. Figure 4 shows the skew-T log-p of the mean of each of the clusters. While the clusters are generally similar near the surface, on average sondes in the shallow moist cluster are colder in the mid-troposphere (up to 700 hPa) than sondes in the deep moist and subsaturated clusters, which may be evidence of the passage of a cold front.

Figure 3

K-means clustering of the vertical moisture profile for radiosondes collected at Bodega Bay during water years 2017 and 2018. We find three distinct centroids, which we classify as “deep moist” (blue), “shallow moist” (black), and “subsaturated” (red). Error bars show the standard deviation of relative humidity in the clusters.

Figure 4

Skew-T of the mean of the (a) deep moist, (b) shallow moist, and (c) subsaturated clusters from Figure 3.

To better understand the physical significance of the different clusters, we consider the timing of the radiosonde launches relative to AR landfall. As an example of this, Figure 5 shows total precipitable water (TPW) from ERA5 averaged between −123.5°E and −122.5°E during the month of February 2017, with the results of the k-mean clustering of the radiosondes launched during this time overlaid on top. From this we can see that the “deep moist” sondes are generally representative of conditions in the AR core, when the TPW at Bodega Bay is highest, while the “shallow moist” profiles were typically taken in the late stages of the AR (though a few were also taken in the early stages before the AR made landfall). The subsaturated profiles commonly occur in between the other two states, and may represent a transition between the deep moist and shallow moist sondes, or a lull in AR conditions. This relationship was true over the entire observation period (not shown). Only two events broke this pattern (January 20, 2017 and March 8, 2018). Both cases featured relatively weak (maximum integrated vapor transport of 474.1 and 406.2 kg m⁻¹ s⁻¹, respectively), short duration (<24 h) events. Comparisons of the timing of the radiosonde launches with the WPC surface archive maps confirms that many of the sondes from the “shallow moist” cluster are associated with the passage of the cold front (WPC, 2019).

Figure 5

ERA5 total precipitable water averaged from −123.5°E and −122.5°E during February 2017. Circles represent the launch time of each radiosonde released from Bodega Bay (38.3°N−123.1°E) during February 2017. Deep moist sondes are red, subsaturated sondes are blue, and shallow moist sondes are black.

As part of this analysis, we also considered clusters based around temperature, wind speed, and wind direction. We found that for temperature and wind speed it was not possible to separate the sondes into well-defined clusters. The exception to this was for wind direction. As with relative humidity, we found three clusters related to vertical profiles of wind direction relating to the life cycle of the AR. During the early and mid stages of the AR, winds were typically southerly at the surface and westerly aloft, transitioned to southerly flow at the surface and southwesterly flow aloft, and finally to southwesterly flow throughout the lower and mid troposphere. These clusters produced similar results, in terms of dust impacts on precipitation, to the relative humidity clusters and are not shown. However, as described in Section 2.3 below, wind direction itself is not part of our model setup; in a more realistic framework, clusters based on wind direction may prove to be an important variable for predicting dust impacts on precipitation.

2.3 Model Description

In this analysis we use the Advanced Research WRF version 3.9.1.1 (Skamarock et al., 2008) run in an idealized two-dimensional setup. Our model domain is 1,200 km long with a horizontal grid spacing of 2 km. The model extends to an altitude of 30 km with 40 vertical eta levels (terrain following). The horizontal length of the domain is necessary to avoid feedback from the lateral boundaries. A 2 km horizontal grid spacing allows us to resolve convection, and the model uses a 20 s time step. The lateral boundaries are open boundaries and the top of the model is a periodic boundary. The model is run with the Thompson Aerosol-Aware microphysics scheme (Thompson & Eidhammer, 2014). The purpose of this study is to focus on the role of microphysics, so other physics packages are not used.

A bell shaped hill is placed in the center of the domain such that

(1)

where h(x) is the height of the topography in km and x is the lateral distance from the center of the domain (km). Figure 1b compares the model topography with a sample transect of topography along the path of an AR. Note that the height of the inland mountain range in California varies from 2 to 4 km (Figure 1a), so 3 km serves as an approximation of the mean height of the Sierras. Each simulation is run for 36 h, with the first 12 h discarded as spin up.

The Thompson Aerosol-Aware microphysics scheme (Thompson & Eidhammer, 2014) is a bulk microphysics scheme which explicitly predicts the mass mixing ratios of cloud water, cloud ice, snow, graupel, and rain as well as the number concentrations of cloud water, cloud ice, and rain. The scheme is an adaption of the previous Thompson microphysics scheme (Thompson et al., 2008) that has been modified to include aerosols acting as CCN and INP. The Thompson scheme is selected over more complex microphysics schemes because it is commonly used in operational forecast models, and in particular is used in West-WRF, a version of the WRF model which has been optimized for forecasting precipitation in the western United States. In order to reduce the computational expense, aerosols are classified as hygroscopic (potential CCN) or non-hygroscopic (potential INP). Hygroscopic aerosols are a combination of sulfates, sea salt, and organic matter. For the purposes of this idealized study, non-hygroscopic aerosols are assumed to be dust. Dust activates into cloud ice following the DeMott et al. (2010) ice nucleation parameterization

(2)

where n_IN,T is the number concentration of activated INP at temperature T, T is the cloud temperature (K), n_INP is the number concentration of INPs, and a, b, c, and d are empirically determined constants, where , , , and d = 0.0033. For the purposes of this theoretical study, we assume that INPs are dust, that is, n_INP = n_dust. Figure 6 shows the relationship between n_IN,T and T for different INP concentrations. In all cases, n_IN,T increases as INP concentration increases and as T decreases. The largest differences between n_IN,T from the different INP scenarios occur at colder temperatures. Supercooled water droplets freeze into ice following the Bigg (1953) scheme, but with the effective temperature modified by the INP concentration following DeMott et al. (2010), such that higher concentrations produce more ice (Thompson & Eidhammer, 2014). Aqueous aerosols freeze into ice crystals following Koop et al. (2000). Secondary ice formation from rime splinters occurs following the Hallet-Mossop process (Hallett & Mossop, 1974; Reisner et al., 1998; Thompson et al., 2008).

Figure 6

Number of activated ice nuclei using the DeMott et al. (2010) parameterization as a function of temperature for different dust concentrations (in cm⁻³).

For the purposes of this experiment, we prescribe background values of CCN to be 300 cm⁻³ (the default concentration in the Thompson scheme). To test the model sensitivity to dust, we consider six different scenarios with dust concentrations of 0.5, 2 , 4, 10, 50, and 100 cm⁻³. Throughout the text we will refer to these scenarios as INP0.5, INP2, INP4, INP10, INP50, and INP100. INP0.5 approximates a climatological average of dust values (Creamean et al., 2014); INP2 and INP4 represent observed values during the CalWater field campaign (Fan et al., 2014). INP10 represents high dust concentrations within a transported dust layer (Fan et al., 2017), and INP50 and INP100 are included to provide the full shape of the power law relationship between dust and ice formation (Section 3), as well as allowing us to span the ranges of results used elsewhere in the literature (Fan et al., 2017). Dust is assumed to have a constant vertical profile at the start of the simulation. Aerosols are removed when they are activated into CCN and INP. While this does not produce a realistic representation of real world dust profiles, it is useful for testing sensitivity to increased dust concentrations in this idealized framework.

We use the radiosondes collected at Bodega Bay (Section 2.1) to force the model at the western lateral boundary. For each dust scenario, we construct a 60-member ensemble by varying the initial conditions at the western lateral boundary using a randomly selected subset of 20 radiosondes (included as supplemental material) from each of the three clusters described in Section 2.2. As described in Section 2.2, the radiosondes were sorted into three clusters based on their vertical profiles of relative humidity. Each sonde provides data on pressure, temperature, relative humidity, wind speed, and wind direction which we use to calculate virtual potential temperature and specific humidity. The variables are then interpolated to 50 m vertical intervals to be input into the idealized WRF model.

3 Dust Sensitivity

As detailed in Section 2.3, we examine the effects of dust on orographic precipitation using the WRF model run with an idealized 2D hill setup. For each dust scenario, we construct an ensemble by forcing the model with 60 of the 245 radiosondes collected at Bodega Bay in 2017–2018. In order to test the precipitation response to dust, we use the low dust (INP0.5) scenario as our control run and perform a series of sensitivity experiments with increased average dust concentrations (n_dust): INP2, INP4, INP10, INP50, and INP100.

The ensemble mean daily average 12–36 h in the simulations) total precipitation (composed of rain, snow, and graupel) in our control scenario (INP0.5) maximizes at 97 mm slightly upwind of the peak of the 3 km hill (Figure 7a). The maximum snowfall (48 mm) occurs at the peak, while the majority of the graupel falls upwind of the peak, with a maximum value of 20 mm. Notably for this study, a considerable portion of the total precipitation occurs as frozen precipitation. Averaged over the upwind slope of the peak (550–600 km from the western boundary) 42% of total precipitation falls as snow, and 18% falls as graupel. In the lee of the peak (600–650 km) 73% of total precipitation occurs as snow and 11% occurs as graupel.

Figure 7

Ensemble mean (60 members) (a) Daily average total precipitation, snow, and graupel in the control scenario (INP0.5) and (b) changes in daily average precipitation, (c) snow, and (d) graupel between the control, and a set of simulations with elevated dust concentrations (INPx-INP0.5). (e) Terrain height is provided for comparison. Gray shaded regions show the location of the mountain.

Increasing dust concentrations (INP ≥ 4) increases the fraction of total precipitation falling as snow and graupel over the peak for all dust scenarios (Figure 7). This shift from rain to frozen precipitation causes total precipitation to decrease on the upwind slope of the mountain, and increase on the downwind slope (Figure 7b). This change is primarily due to the increase in the amount of precipitation falling as snow at higher dust concentrations. The net effect on precipitation over the mountain is small, but the increased snow/rain ratio advects precipitation toward the lee side of the mountain. This displacement is sometimes referred to as a “spillover effect,” and occurs as a result of the slower fall speed of snow compared to rain (Colle, 2004; Colle et al., 2005; Colle & Mass, 2000; Colle & Zeng, 2004; Morales et al., 2018; Wallmann & Milne, 2007). The increase in the percentage of precipitation falling as snow and graupel is driven by increases in ice water path (IWP), particularly upwind of the peak (Figure 8c). This increase comes at the expense of LWP, which decreases by a similar amount over the same region (Figure 8b). Additionally, the increase in graupel upwind of the peak, on the order of 0.7 mm (5%; INP2) to 2.0 mm (12%; INP100) averaged from 550 to 600 km, is evidence that there is an increase in riming processes due to increased dust concentrations. This suggests that in the ensemble mean, dust may be enhancing the seeder-feeder mechanism, but that the overall effect on precipitation is small relative to the orographic forcing of the mountain (on the order of 0.1% for INP2 to 0.4% for INP100 averaged from 550 to 650 km).

Figure 8

(a) Daily average ensemble mean liquid water path (LWP, green) and ice water path (IWP, blue) in the control scenario. (b) Changes in daily average LWP and (c) IWP between the control, and a set of simulations with elevated dust concentrations (INPx-INP0.5). Gray shaded region shows the location of the mountain.

Fan et al. (2014, 2017) demonstrate that the impacts of dust can vary significantly depending on the characteristics of the storm, and so we sort the simulations based on the clustering of the input sondes described in Section 2.2. By design, the 60 input sondes were randomly selected so that there are 20 sondes from each cluster. Figure 9 shows the average total precipitation, snow, and graupel in the control run for each of the three clusters. Unsurprisingly, the average total precipitation, snowfall, and graupel are greatest for the deep moist simulations, and least in the subsaturated case. Comparing the deep moist and shallow moist simulations, the overall precipitation totals are similar, but the percentage of precipitation falling as snow is smaller in the shallow moist case due to the lower moisture availability above the freezing level (Figure 4b). The deep moist cluster has stronger updrafts upwind of the mountain, with a mean vertical velocity of 1.19 m s⁻¹ averaged from 550 to 600 km and from the surface to 5 km (Figure 10a). In contrast, the updrafts upwind of the mountain in the shallow moist (mean vertical velocity of 0.77 m s⁻¹, Figure 10b) and subsaturated (mean vertical velocity of 0.66 m s⁻¹, Figure 10c) cases are relatively weak. The cloud layer in the deep moist cluster extends to heights of 12 km, even before being lifted orographically (Figures 11a and 11b). In contrast, the shallow moist cluster's cloud layer is capped at around 5 km before being lifted (Figures 11c and 11d), while the subsaturated cluster has a low cloud (also capped around 5 km), as well as a high ice cloud in the upper troposphere (up to 15 km; Figures 11e and 11f). The shallow moist cluster is a purely warm cloud until it is orographically lifted and begins to form ice (Figures 11c and 11d). As a result of their weaker convection, the shallow moist (Figure 12b) and subsaturated (Figure 12c) clusters have significantly more supercooled water available (0.33 and 0.29 g kg⁻¹ averaged from 500 to 600 km from the western boundary and from the surface to 5 km) than the deep moist cluster (0.19 g kg⁻¹, Figure 12a), which already has significant ice formation in the low dust simulation.

Figure 9

As in Figure 7a, but with the ensemble members split into the (a) deep moist, (b) shallow moist, and (c) subsaturated clusters shown in Figure 3.

Figure 10

Ensemble average vertical velocities (m/s) in the (a) deep moist, (b) shallow moist, and (c) subsaturated clusters.

Figure 11

Vertical distribution of cloud ice mixing ratio (q_i, blue) and cloud water mixing ratio (q_c, green) in the (a and b) deep moist, (c and d) shallow moist, and (e and f) subsaturated clusters at 200 km (left) and 550 km (right) from the western boundary. Cloud ice has been multiplied by 100 so that it can be plotted on the same scale as cloud water. The black line shows n_IN,T. All panels are for the control scenario (INP0.5).

Figure 12

Vertical distribution of supercooled water droplets (cloud water at T < 273.15 K) in the (a) deep moist, (b) shallow moist, and (c) subsaturated clusters for the control scenario (INP0.5).

As seen in the ensemble average (Figures 7b, 7c, and 7d), increasing the dust concentration leads to increases in the snowfall over the mountain (Figures 13d, 13e, and 13f), increases in graupel upwind of the peak (Figures 13g, 13h, and 13i), and decreases in total precipitation upwind of the peak coupled with increases in total precipitation in the lee of the peak (Figure 13a, 13b, and 13c) in all clusters. The changes in total precipitation upwind of the peak are small relative to the precipitation in the control (decreases on the order of 1 mm or less, representing changes of 1% or less). Downwind of the peak, the increases in total precipitation are on the order of 0.2 mm (1%; INP2) to 1 mm (5%; INP100) in each of the clusters. The total change in precipitation averaged over the peak (550–650 km) is not significantly different from zero for any cluster or INP concentration (Figures 14a, 14b, and 14c), where significance is determined using a student t-test with 95% confidence.

Figure 13

Changes in total precipitation (ΔTotal, top), snow (ΔSnow, middle) and graupel (ΔGraupel, bottom) for the (a, d, and g) deep moist, (b, e, and h) shallow moist, and (c, f, and i) subsaturated clusters shown in Figure 3.

Figure 14

Boxplots show the ensemble spread of the change in precipitation, snow and graupel averaged over the peak (550–650 km) for the deep moist (a, d, and g), shallow moist (b, e, and h), and subsaturated clusters (c, f, and i). Circles depict the ensemble means. Filled circles indicated that the mean is significantly different from 0 at the 95% confidence level using a student t-test. Red crosses denote outliers.

There are notable differences in the relative and absolute magnitudes of the modeled changes in frozen precipitation. In the subsaturated and deep moist cases, changes in snowfall range from near zero (INP2), to increases of 4 mm (INP50 and INP100; Figures 13d and 13f). In contrast, in the shallow moist case, there are clear increases in snowfall, especially at lower dust concentrations (2–3 mm at INP2 and INP4, up to 6 mm at IN100; Figure 13e). In relative terms, the changes in snowfall also represent a much larger percentage increase in the shallow moist case: 9%–25% (INP2 to INP100) over the peak, compared with 6%–18% in the subsaturated case and 2%–10% in the deep moist case. Averaged over the peak, we find that the mean changes in snow are significant for all INP concentrations in the shallow cluster (Figure 14e), and for INP4 and INP50 for the subsaturated cluster (Figure 14f). The deep moist cluster has two outlier cases that were extremely sensitive to increased INPs (not shown for INP100), but the ensemble mean did not differ significantly from zero (Figure 14d).

When considering graupel on the other hand, the shallow moist case shows the smallest changes in both the absolute and relative sense. Averaged over the upwind slope of the peak (550–600 km from the western boundary), graupel increased by 0.16 mm (INP2) to 0.52 mm (INP100), with maximum increases of up to 2.3 mm (Figure 13h). These changes represent 0.5%–3.0% increases in graupel. The absolute changes in graupel are similar in the subsaturated and deep moist cases for the higher dust concentrations (INP10 through INP100), on the order of 1–3 mm, but at the lower concentrations (INP2 and INP4), the changes in graupel are larger in the deep moist case (0.8–1.0 mm) compared with the subsaturated case (0.3–0.5 mm; Figures 13g and 13i). For INP10 through INP100, the absolute changes in graupel in the subsaturated cluster represent a much higher relative change ranging from 10% to 20% averaged over the upwind slope of the peak (with maximum values as high as 30%). In contrast the changes in the deep moist case represent 5%–10% increases in graupel. The changes in graupel over the peak are significant at higher INP concentrations (INP10 through INP100) in the subsaturated cluster, though generally not significant in the other clusters (Figures 14g, 14h, and 14i). It is noteworthy that in the shallow moist and deep moist clusters, the variance in snow and graupel generally increases as the INP concentration increases, indicating that some cases within these clusters are highly sensitive to INPs, while others change relatively little.

These changes in precipitation can be traced to changes in LWP and IWP, shown in Figure 15. The largest and most significant changes in LWP and IWP occur in the shallow moist case. This is driven by the relatively large amount of supercooled water in the low dust case being converted to snow. The smallest changes occur in the deep moist simulations, likely due to the fact that the input profiles are already at or near saturation through the mid-troposphere, and the relative lack of supercooled water in the low dust case. These changes follow a power-law relationship as a function of dust concentration (Figure 16), due to the functional relationship between n_dust and n_IN,T (Equation 2). Changes in LWP are nearly equal and opposite to changes in IWP, indicating that the growth of ice is coming primarily at the expense of liquid water, rather than water vapor. LWP and IWP are most sensitive to dust at lower concentrations (INP ≤ 10).

Figure 15

As in Figure 8, but with the ensemble members split into the (a, d, and g) deep moist, (b, e, and h) shallow moist, and (c, f, and i) subsaturated clusters.

Figure 16

Changes (INPx-INP0.5) in IWP (a) and LWP (b) averaged over the peak (550–650 km) as a function of dust concentration for the deep moist cluster (red), shallow moist cluster (black), and subsaturated cluster (blue).

4 Model Sensitivity

To assess the robustness of our results, we perform further analyses examining the sensitivity to different modeling choices. In this section, we consider the effects of different ice nucleation parameterizations, model resolution, and the addition of a second mountain, analogous to the coastal range in California. Due to computational constraints, we perform these sensitivity tests on a subset of the 60 ensembles members used in the main body of, the paper, selecting three radiosondes from each cluster.

The results presented in Section 3 use the DeMott et al. (2010) ice nucleation parameterization (Equation 2), which was derived using measurements of INPs from a series of observations mostly made over the Western US. Here we present a comparison with the DeMott et al. (2015) ice nucleation parameterization:

(3)

where n_IN,T is the number concentration of activated INP at temperature T, T is the cloud temperature (K), n_INP is the number concentration of INPs, and a, b, c, and d are empirically determined constants, and cf is a calibration factor. Here, a = 0, b = 1.25, c = 0.46, and d = −11.6. This parameterization was derived from laboratory based studies and is designed to provide a global approximation of dust effects on ice nucleation. We use a calibration factor of 3, as derived in DeMott et al. (2015) for atmospheric data. In a case study, this was also shown to provide good agreement with the Niemand et al. (2012) parameterization in a Saharan dust layer, although more work would be required to determine the relationship between these two parameterizations in a broader context (DeMott et al., 2015). At low dust concentrations Equations 2 and 3 produce similar results, but n_IN,T in Equation 3 is much more sensitive to higher values of n_INP, representing the higher ice nucleation activity of dust relative to other INPs.

In the control simulation (INP0.5), the parameterization had very little effect on precipitation in the cases tested (Figures 17a and 17b) as expected. At higher dust concentrations, the DeMott et al. (2015) parameterization lead to more ice being formed relative to DeMott et al. (2010). Comparing Figures 18 and 19, we see larger increases in snow and graupel using the DeMott et al. (2015) parameterization and a more prominent spillover effect. Averaged on the upwind slope of the peak (550–600 km), total precipitation decreases by 1.33 mm (INP2) to 2.65 mm (INP100) using the DeMott et al. (2010) parameterization, snow increases by 2.36 mm (INP2) to 8.37 mm (INP100), and graupel increases by 0.39 mm (INP2) to 3.36 mm (INP100). In contrast, using the DeMott et al. (2015) parameterization, precipitation decreases by 1.39 mm (INP2) to 4.37 mm (INP100), snow increases by 4.42 mm (INP2) to 17.75 mm (INP100), and graupel increases by 0.91 mm (INP2) to 4.44 mm (INP100). The differences between parameterizations are most prominent at high dust concentrations, but even at INP2, the changes in frozen precipitation (snow and graupel) are approximately doubled. The changes in precipitation agree qualitatively between the two parameterizations, but this suggests that the results presented in Section 3 may represent a lower bound on dust impacts on orographic precipitation over the western U.S.

Figure 17

Ensemble mean (nine members) daily average total precipitation, snow, and graupel in the control scenario (INP0.5) using the (a) DeMott et al. (2010) ice nucleation parameterization (b) DeMott et al. (2015) ice nucleation parameterization, (c) increased horizontal resolution (1 km), and (d) a second small hill (500 m) analogous to the California coastal range.

Figure 18

Ensemble mean (nine members) change in (a) total precipitation, (b) snow, and (c) graupel (INPx-INP0.5) using the DeMott et al. (2010) ice nucleation parameterization.

Figure 19

Ensemble mean (nine members) change in (a) total precipitation, (b) snow, and (c) graupel (INPx-INP0.5) using the DeMott et al. (2015) ice nucleation parameterization.

Similarly, we tested the effects of model grid spacing by re-running the nine simulations described above, but with the horizontal grid spacing halved to 1 km. The change in grid spacing had minimal effects on the control simulations (Figure 17c). Averaged over the upwind slope of the mountain (550–600 km), total precipitation decreased by −0.86 mm (INP2) to −3.66 mm (INP100). Snow increased by 3.19 mm (INP2) to 14.92 mm (INP100), and graupel increased by 1.14 mm (INP2) to 4.66 mm (INP100). Compared with the 2 km grid spacing simulations (Figure 19), these simulations have smaller changes in precipitation and snow, while graupel is slightly more sensitive to dust (Figure 20).

Figure 20

Ensemble mean (nine members) change in (a) total precipitation, (b) snow, and (c) graupel (INPx-INP0.5) using the DeMott et al. (2015) ice nucleation parameterization, and with the horizontal grid spacing halved to 1 km.

Finally, while our goal in this paper has been to present results that are generalizable beyond the US West Coast, the West Coast does have important terrain features that may have an effect on our results. To test the robustness of our results, we performed an experiment where we added a coastal mountain range, with a height of 500 m, centered at 400 km from the western boundary. The addition of the small hill produced a secondary peak in total precipitation centered over the hill that is composed entirely of rain (as opposed to snow or graupel; Figures 17a and 17d). This Double Hill test had a relatively small impact, except at high dust concentrations (INP50-100, compare Figures 19 and 21). Total precipitation over the upwind slope of the 3 km peak (averaged from 550 to 600 km) decreases by 1.15 mm (INP2) to 3.21 mm (INP100) and snow increases by 2.14 mm (INP2) to 11.03 mm (INP100; Figure 21). Compared with the changes in the single hill simulations (Figure 19), this represents a slight decrease in the dust sensitivity of snow and total precipitation. The increase in graupel falling on the upwind slope of the 3 km peak was similar to the single hill simulations in the control simulations (0.83 mm for INP2), but at high dust concentrations, graupel was more sensitive to dust under the two hill scenario (6.32 mm at INP100). An important caveat to this test is that the number of vertical levels (40) used here is relatively coarse and simulations run with more vertical levels might be more sensitive to the addition of the smaller hill.

Figure 21

Ensemble mean (nine members) change in (a) total precipitation, (b) snow, and (c) graupel (INPx-INP0.5) using the DeMott et al. (2015) ice nucleation parameterization, with a second hill (500 m) added centered 400 km from the western boundary.

5 Discussion

Overall, the effects of dust on total precipitation were relatively small (generally ≤ 1.5% upwind of the peak, Figure 7b), but we did find that dust had a large effect on precipitation type (Figures 7c and 7d), leading to increases in both snow and graupel (as much as 10% upwind of the peak at the highest dust concentrations) in our idealized simulations. The ability to accurately forecast the snow/rain ratio during landfalling ARs has important implications for water resource management (Dettinger et al., 2011; Ralph et al., 2019). Additionally, the snow/rain ratio is important for understanding flood risks both during and after events. When more of the precipitation falls as rain, it will increase the risk of flooding during the AR (Lundquist et al., 2008), although at the same time, a higher ratio falling as snow could create antecedent conditions that would lead to greater flood risks during subsequent events (Kattelmann, 1997). The increases in precipitation on the lee side of the peak, sometimes referred to as a “spillover” effect also provide an important source of water for areas to the east of the mountain.

In general, the relationship between dust concentration and LWP and IWP follows a power law relationship, and is most sensitive at lower concentration levels (INP <10 cm⁻³, Figure 16), resulting in a non-linear precipitation response (Figures 7b–7d). This suggests that at higher dust concentrations, moisture availability becomes the determining factor for ice formation, rather than temperature. We found that the sondes that we classified as “shallow moist” were most sensitive to changes in dust concentrations (Figures 13 and 16). In these cases, the environment was on average colder than other sondes, with a moist layer near the surface that is capped in the lower troposphere. Unlike the deep moist sondes, which tended to be saturated throughout the mid-troposphere, or the subsaturated sondes which are below saturation throughout most of the troposphere, these sondes only become subsaturated near the freezing level. As such, adding dust (which effectively increases the temperatures at which ice can form in the model), will have a large impact on the amount of moisture that is available for ice nucleation. The shallow moist sondes represent conditions on the periphery of ARs. About 11 of the 20 sondes that were included in the shallow moist cluster occurred on or after the passage of the cold front at Bodega Bay (not shown), indicating that precipitation occurring along with the cold front may be especially responsive to dust. In addition, previous research has indicated that the cold sector of a storm is the region where dust is most likely to be present (Creamean et al., 2013). While the bulk of precipitation during an AR typically falls prior to the passage of the cold front, narrow cold frontal rainbands produce short duration intense precipitation that has been associated with hazardous debris flow (Oakley et al., 2017). The potential role of atmospheric dust in contributing to these brief intense precipitation events should be evaluated in future studies.

Previous modeling and observational studies have found that in some cases increased dust concentrations can lead to increases in total precipitation (rain and snow) via the seeder-feeder mechanism (Ault et al., 2011; Creamean et al., 2013; Fan et al., 2014, 2017). Our model is unable to reproduce this result in the ensemble mean. Although increasing dust leads to increasing snowfall over the mountain (Figure 7c), total precipitation decreases upwind of the peak (Figure 7b). The only increases in total precipitation occurred on the downwind slope of the peak, where most of the precipitation fell as snow in the control simulation (Figure 7b). However, a few individual ensemble members did produce increases in total precipitation. Figure 22 was the first radiosonde collected during a January 21–22, 2018 AR event, and was classified as subsaturated in our clustering. This sonde was relatively cold in the lower atmosphere and has a pronounced dry layer from 900 to 750 hPa. Notably, this radiosonde has the most pronounced dry layer of all the radiosondes collected during the 2017–2018 FIRO campaign. This dry layer is an important element of a typical seeder-feeder environment because it indicates that the high cloud is decoupled from the low cloud (Schneider & Moneypenny, 2002; Thompson et al., 2004). In this case, the initial conditions were cold enough that the model produced snow upwind of the mountain in the control simulation (Figure 23a). Figures 24a and 24b shows the vertical distribution of cloud ice, cloud water, snow and graupel in the low dust simulation at 200 km distance. Ice is concentrated in the layer between 5 and 10 km. Below 5 km, ice develops into snow and graupel and begins to precipitate out. As shown in Figures 23b and 23c, when dust is added to the simulation, it increases snow on the upwind slope of the mountain (400–600 km) by 4.37–6.10 mm (INP2.0–INP100) and total precipitation by 4.40–6.57 mm (INP2.0–INP100). Graupel goes from nearly non-existent in the low dust concentrations (control, INP2, and INP4) to 1–2 mm in the higher dust concentrations (INP10 - INP100, Figure 23d). Focusing on INP10, there is a large increase in cloud ice in the mid troposphere, and a corresponding increase in snow and graupel (Figure 24c and 24d). However, in this case, there is also an increase in cloud water near the surface. We hypothesize that frozen precipitation (snow and graupel) falling through the dry layer near the surface may have melted/sublimated, providing a source of moisture and allowing for the increase in cloud water. This process resembles the seeder-feeder mechanism, wherein precipitation in the low cloud is fed by snow and ice falling from a higher cloud (Creamean et al., 2013). This supports the interpretation that the seeder-feeder mechanism is most important during the beginning and end of the event, which is not necessarily well represented by the FIRO radiosondes as the project focused on peak AR intensity.

Figure 22

Skew-T for the radiosonde launched from Bodega Bay on January 21, 2018 at 18Z.

Figure 23

(a) Daily average precipitation, snow and graupel in the control scenario forced by the sonde in Figure 22. Changes in daily average (b) precipitation, (c) snow, and (d) graupel (as in Figure 7b) for the single ensemble member.

Figure 24

(a and c) Vertical distribution of cloud ice (blue) and cloud water (green) at 200 km for the radiosonde launched from Bodega Bay on January 21, 2018 at 18Z. Cloud ice is multiplied by 100 so that it can be plotted on the same scale as cloud droplets. (b and d) Vertical distribution of total frozen water (snow, graupel, and ice; blue) and total liquid water (rain and cloud water; green). Frozen (liquid) water is predominantly snow (rain). The black line shows n_IN,T. The top plots (a) (b) show INP0.5. The bottom plots (c) (d) show INP10.

6 Conclusions

In this study, we use a theoretical modeling framework to test the sensitivity of orographic precipitation to heightened dust concentrations under a broad range of AR initial conditions. We found that increasing dust increased the percentage of total precipitation that was falling as frozen precipitation (snow and graupel). The slower fall speeds of snow relative to liquid rain produced a spillover effect, where total precipitation decreased upwind of the peak and increased in the lee of the peak. The modeled precipitation was most sensitive to dust when it was initiated with “shallow moist” conditions, which primarily occur at the beginning and end of AR events. In general, the modeled sensitivity to dust followed a power law relationship, as predicted by the DeMott et al. (2010) ice nucleation scheme.

In order to test the robustness of our results, we ran a smaller ensemble and tested the effects of using a different ice nucleation parameterization, increasing the model resolution, and adding a second, smaller hill similar to the California coastal range. We found that using the DeMott et al. (2015) ice nucleation parameterization led to the model being far more sensitive to changes in dust. In particular, the increases in snow caused by dust approximately doubled compared with the DeMott et al. (2010) parameterization. Increasing the model resolution had a smaller impact, but did lead to a small increase (decrease) in the sensitivity of graupel (snow) at high dust concentrations, potentially due to stronger updrafts at the higher resolutions (not shown). Similarly, adding a second 500 m hill to the model also lead to an increase (decrease) in the sensitivity of graupel (snow) at high dust concentrations.

As we have shown here, dust is important for determining the snow/rain ratio during ARs, particularly at the early and late stages of the event, and in individual cases may have a large impact on overall precipitation. However, further research is needed to fully understand the effects of dust on orographic precipitation during landfalling ARs. In this work, we assumed a constant vertical profile of dust, but in the real atmosphere dust is transported across the Pacific in discrete layers, and we expect the altitude of the dust layer to affect the precipitation response (Ault et al., 2011; Creamean et al., 2013). This study neglects the role of large scale dynamics, in particular the Sierra barrier jet, which is expected to contribute to the seeder-feeder mechanism by dissociating the upper level seeder cloud and the lower level feeder cloud. In addition, other physics, such as cloud-radiation interactions were not considered and may provide another avenue by which dust can affect orographic precipitation. Further studies will be needed to test the robustness of these results to different model configurations, such as using a more computational expensive spectral bin microphysics scheme, rather than the Thompson Aerosol Aware microphysics. Finally, in order to better validate the results of this work we will need to obtain collocated observations of vertical profiles of dust (and other INPs), temperature, humidity, and hydrometeors during landfalling ARs.

Acknowledgments

The authors would like to thank Kara Voss and Andrew Martin for their contributions to this project. This work was funded by the California Department of Water Resources contract 4600010378, Task Order OSCOP215 and the Army Corps of Engineers USACE (CESU) W912HZ-15-0019.

Open Research

Data Availability Statement

All FIRO radiosondes are available as part of the UC San Diego Library Digital Collections at https://doi.org/10.6075/J09P31S0 (Ralph et al., 2021).