Throughout the Gulf of Mexico (GoM), the Loop Current (LC) drives essentially all mesoscale oceanographic variability (Lugo-Fernández et al., 2016; Hamilton et al., 2016). The LC enters the GoM through the Yucatán Strait from the Caribbean Sea as a warm and salty jet (Sturges and Leben, 2000), and on irregular intervals of approximately 3–18 months, it pinches off an anticyclonic Loop Current Eddy (LCE) that moves separately to the north and west (Leben, 2005). The LC and LCEs carry strong currents that can be particularly disruptive to offshore marine operations along the northern Gulf (Kantha, 2014), and additionally, if a hurricane passes over the LC or an LCE, the excess heat in their near-surface waters causes intensification (Le Hénaff et al., 2021). These several effects prioritize an understanding of the dynamics of the LC system to promote public safety.

A sequence of historical efforts have observationally, theoretically, and numerically investigated mechanisms that extend the LC farther into the GoM together with the processes by which LCEs are shed (Pichevin and Nof, 1997; Hurlburt and Thompson, 1982; Schmitz, 2005; Oey et al., 2005). Much recent focus has been on near-surface observations and dynamics, seeking to develop improved forecasting ability for LCE separation (Dukhovskoy et al., 2023; Ernst et al., 2023). In contrast, knowledge of the processes governing the deep circulation remains elusive, though it is clear that observed deep near-bottom eddy variability can be strong (∼ 0.1 − 0.4 m s⁻¹) in much of the GoM and almost never quiescent (Donohue et al., 2016b; Johnson et al., 2022; Furey et al., 2018; Pérez-Brunius et al., 2018). Numerical models have typically underestimated the deep eddy kinetic energy by a factor of three or more (Rosburg et al., 2016; Morey et al., 2020).

Observations have illustrated ways that the upper and deep GoM interact. To leading order, the circulation in the GoM is characterized by two vertical modes (Hamilton, 2009; Hamilton et al., 2016; Pérez-Brunius et al., 2018), with one describing the strong upper baroclinic currents associated with the LC and LCEs, and the second mode being nearly depth-independent from sea surface to the bottom (Inoue et al., 2008; Welsh et al., 2009; Hamilton et al., 2011). Together these two modes were found to account for ∼ 98% of the variance of subinertial currents in the central GoM (Donohue et al., 2016b). Mutual upper-deep coupling mechanisms are the focus of this paper.

Observing the characteristic signatures of upper-deep coupling requires measurements of currents at multiple levels spanning the upper and deep water column, and is facilitated by mapping arrays of the pressure stream function fields in the near surface and at depth. Such datasets have generally been confined to latitudes north of 25.5°N (Hamilton et al., 2015). Additionally, numerical modeling efforts have not focused south of this latitude; however Hurlburt (1986) and Oey (2008) suggested that the region immediately north of Campeche Bank, where the LC enters deeper water (Figure 1), is an important area for generation of deep eddies. Large mean-to-eddy energy conversion rates appear there along the western edge of the Loop Current as the current moves off the Campeche Bank into the deep topography of the Gulf.

Complementing that earlier work, this paper highlights upper-deep interactions further south, in a region that has not been examined before. A 6-year free-running numerical simulation using a regional configuration of the Navy Coastal Ocean Model (NCOM), to be described in Section 2, shows the importance of deep cyclones to LCE formation, particularly in the narrowly confined deep region south of 25°N, circumscribed by the Yucatán Strait to the south and the steep lateral sidewalls of the Campeche Bank and the Florida Shelf to the west and east, respectively. Section 3 discusses deep eddies and vertical stretching. The model run highlights four southern separation events, as presented in Section 4. They illustrate ways that deep cyclones develop and strengthen from baroclinic joint development as well as from lower-layer stretching when the LC moves away from the steep channel sidewalls. Section 5 summarizes results.

We use the Navy Coastal Ocean Model (NCOM) Version 4.3 (Martin et al., 2009), a primitive equation, hydrostatic model with the Boussinesq approximation. This model and configuration underpin an ensemble forecast system for the Gulf of Mexico Thoppil et al. (2025). Therefore, the model architecture, basic state, and performance have been tested and verified. We have selected case studies that are representative of LCE separations in the Deep Southeast Channel. The model uses a staggered Arakawa-C grid that extends from 18°N to 31°N and 98°W to 79°W with a 3 km horizontal resolution on a spherical projection. In the vertical, there is a 35 layer terrain-following (free-σ) grid in the upper 500 m with bottom cells “shaved” as needed to match bathymetry, and below that are 15 z-levels. Within the Loop Current region, the surface and bottom cells are approximately 0.5 and 500 m thick, respectively.

Vertical mixing is parameterized using the Mellor-Yamada Level 2 closure model (Mellor and Yamada, 1974, 1982). A logarithmic bottom boundary layer is assumed, and bottom drag is computed with a spatially constant roughness length of 1 cm. All advection schemes are 2nd order and no tidal forcing is prescribed. Open boundary conditions are specified by the global 1/12° 41-layer Hybrid Coordinate Ocean Model (HYCOM) with the Naval Research Lab Coupled Ocean Data Assimilation (NCODA) System, together known as the Global Ocean Forecasting System (GOFS) version 3.1 (Metzger et al., 2014). The model was initiated from the year 2018 after a deep salinity offset had been corrected in GOFS 3.1. Deep temperature and salinity fields were relaxed to the U.S. Navy Generalized Digital Environment Model (GDEM) version 3.0 (Carnes et al., 2010). Meteorological forcing is from the Navy Global Environmental Model (Hogan et al., 2014). Model output (currents and sea surface height) were 4-day low-pass filtered for analysis. The model integration is run without data assimilation to forecast Loop Current eddies for comparison with observations. Instead our purpose is to illustrate the free-running dynamics that occur within the whole water column around the LC during its naturally occurring extension, detachment/reattachment, and separation cycle.

To represent the deep eddies, we extract the deep reference pressure (scaled hydrostatically to its equivalent sea surface height) after decomposing the total model sea surface height anomaly (η_tot) as the sum of deep reference (η_ref) and steric (η_steric) components:

where η_steric is equal to the geopotential height anomaly, ϕ, divided by the acceleration of gravity, g. To compute ϕ, we integrate the specific volume anomaly, δ=(1ρ(S,T,P)−1ρ(S0,T0,P)) from the sea surface (P₁ = 0 dbar) to a constant reference pressure of P₂ = 2023 dbar (a convenient reference pressure in the model):

Here P is pressure, ρ is density, S is salinity, T is temperature, and S₀ and T₀ are reference values of 35 psu and 0°C, respectively. Hence, η_ref is the perturbation height of the reference pressure surface, calculated as the difference between η_tot and ηsteric=ϕ/g integrated in the following case-studies from the sea surface down to 2023 dbar.

Three sources of observations were obtained with which to compare model output. In situ observations consist of ∼ 4000 CTD and Argo float profiles, from approximately 82.9°W and 23.3°N to 87.5°W and 28.6°N in the eastern GoM, and from 87.5°W and 23.9°N to 96.1°W and 28.6°N in the western GoM. Besides the majority of profiles shallower than 2000 m that were provided by the Argo float program between 2010 and 2021, the bulk of the deep data came from shipboard CTDs in the Exploratory Study of Deepwater Currents in the Gulf of Mexico (Donohue et al., 2006). These were supplemented by hydrocasts, National Oceanic and Atmospheric Administration (NOAA) casts during the Deepwater Horizon oil spill, and observations from the Center for Scientific Research and Higher Education at Ensenada (CICESE; Donohue et al., 2016a). For comparison, both model output and hydrographic profiles were averaged in two separate and narrow ranges of geostrophic stream function (ϕ in Equation 2), to distinguish between inside and outside the warm core of the LC. Profiles of salinity are shown in Figure 2. Outside of the LC, the modeled salinity demonstrates a small fresh bias, while inside the LC, the model is notably fresh by approximately 0.3 psu near the ∼200 m depth Subtropical Underwater (SUW) salinity maximum.

We have evaluated altimeter-derived SSH standard deviation (not shown) for the years 2018-2023. The magnitudes are comparable, and maximum standard deviations are near 0.28 m and are associated with the LC retreating and extending into the Gulf. The model has smaller SSH standard deviation northwest of 25°N compared to the altimeter data, which simply corresponds to the present set of cases considered here where southern separations are favored.

Leben (2005) showed that the separation period was influenced by LC retreat, with a bimodal distribution: when the LC retreats south of 25°N, the average separation period for subsequent separations was 16.2 months – longer than the case when the LC retreats north of 25°N. The present model runs for 6 years and has 6 separations, four of which are detailed in this work. From simulated years 2019 to 2022, four sequential separations occurred where the LC retreated south of 25°N. These separation intervals are on par with the Leben (2005) altimeter-based observations that ranged from 1 to 18 months.

To compare the eddy kinetic energy (EKE) of deep currents, gridded float data from 1500–2500 m were obtained from the Bureau of Ocean Energy Management (BOEM). For the gridded data, EKE was estimated as half of the sum of the variance along the major and minor current variance ellipse axes (Pérez-Brunius et al., 2018). Time-averaged EKE for the gridded and model data is shown in Figures 2C, D. The observed deep EKE in the southeastern GoM is approximately 2–3 times larger than that computed from the model. This low bias is similar to underestimates of deep EKE observed in other models (Rosburg et al., 2016; Morey et al., 2020). Nevertheless, we will demonstrate that characteristics of baroclinic instability processes are realistically depicted in the present model.

The time-averaged deep pressure field (η_ref) from the numerical simulation shows that the prominent feature in the southeastern GoM is a deep cyclone that spans the width of the local sub-basin (Figure 3A). We refer to this deep basin south of 25°N, circumscribed by the Yucatán Strait to the south and the steep lateral sidewalls of the Campeche Bank and the Florida Shelf to the west and east, respectively, as the Deep Southeast Channel, or DSC (Figure 1). From six years of model output, the average deep reference height within the deep cyclone shown in Figure 3A is -0.0076 m, and its amplitude ranges from ∼ +0.06 m to ∼ −0.10 m during eddy shedding events (Figure 3B).

All four eddy shedding events that occurred during the model run are presented as Case Studies in Section 4. LC interactions with deep cyclones and anticyclones are important during these events. Examination of Figure 3B prompts two remarks. First, several additional deep pressure lows in the DSC occurred without separation of an LCE. Some deep lows (more examples than illustrated here) occurred during intervals of a retracted LC, suggesting a deep cyclone may tend to block LC extension, such as illustrated in Case 3. Also, the deep cyclone does not initially have to be strong, but can develop to nonetheless influence the necking down or shedding process, as will be shown in Case 4.

Prior numerical studies have noted cyclonic features in the southeast GoM (Oey, 2008; Morey et al., 2020), yet the mechanisms of formation for these deep cyclones has not been explored and deserves further observational and numerical investigation. Based on the numerical model results presented here, we argue below that deep cyclones in the DSC region are important in regards to LCEs because of their contributions to advection and vorticity, facilitating necking down or pinching off of an LCE. In addition, they may prevent previously detached eddies from reattaching (Schmitz Jr, 2005; Chérubin et al., 2006; Shay et al., 2011).

Importantly, though the η_ref contribution to sea surface height is relatively weaker than the range of η_tot, the signatures, gradients, and currents dominate the sub-pycnocline layers below ∼ 1200 m. While we will refer to them as deep eddies and deep cyclones, one should bear in mind that they contribute a uniform vertical structure throughout the water column. In particular, their reference velocities add components that may cross normal to the upper layer baroclinic currents (Donohue et al., 2016b), and hence in the upper water column the reference currents are predominantly responsible for mesoscale cross-frontal motions.

The mechanism by which the upper and lower layers influence each other is through vertical stretching exerted by one layer upon the other layer, as in the following two manners: (i) when the upper baroclinic front shifts laterally, the sloped isopycnals enable vertical stretching of the weakly stratified lower water column; (ii) additionally, when the deep reference current has a component of flow crossing the upper baroclinic jet structure, its depth-independent component flows along sloped isopycnal levels z_ρ. This produces vertical velocity w=u·∇(zρ) in the upper layer (Lindstrom et al., 1997), and consequently, the upper layer waters are vertically stretched or squashed, as shown in Howden (2000). Both aspects of upper-deep coupling work jointly in the well-known process of baroclinic instability, which develops when there is a favorable vertical offset between perturbations in the upper and deep eddy variability. The characteristic signature is a phase tilt with deep cyclones/anticyclones leading the upper jet troughs/crests along the direction of upper jet flow by roughly a quarter wavelength (Watts et al., 1995; Sutyrin et al., 2001). If deep flow components cross sloped bottom topography, this contributes strong additional vertical stretching. A textbook illustration can be seen in Cushman-Roisin and Beckers (2011), and observational studies include Savidge and Bane (1999); Chereskin et al. (2009); Donohue et al. (2016b), and Furey et al. (2018).

To illustrate the dynamical relationship between relative vorticity in the lower layer and the LC pycnocline depth, we begin with conservation of potential vorticity (Q), and follow procedures from Vallis (2017), allowing us to write (Equations 3–5):

and

where ζi is relative vorticity in layer i, f is the Coriolis frequency, and hi is the thickness of layer i. For convenience, we proceed by omitting the subscript i, and focus on the deep layer. The Coriolis frequency varies with latitude, approximately linearly as f=f0+βy, where β=df/dy, and y is the latitudinal dimension. Decomposing h into mean layer thickness H and perturbation h′, we write h=H(1+h′H). Assuming both |ζ|≪f0 and |βy|≪f0, as well as |h′|≪H, and making use of the binomial expansion gives:

f0H is constant, and in our examples the small meridional excursions make βyf0≪h′H (respectively, these two quantities are of order 1E-2 and 1E-1). Hence following a water parcel the quantity q≈ζf0−h′H is materially conserved, implying that changes in relative vorticity correspond to changes in layer thickness as:

In the upper layer, h′ changes with pycnocline depth when a parcel has a cross-frontal component of flow. In the lower layer, h′ changes with both pycnocline depth and with bottom topography, whereby either influence can produce squashing or stretching of the deep layer.

Importantly, within the DSC, bracketed by nearly vertical side walls (Figure 4), the deep stretching is locally constrained, and the deep eddies so produced cannot translate away. An idealized schematic of this latter process is illustrated in Figure 5, showing the development of a deep cyclone in the stretching lower layer beneath the departing LC pycnocline. The effectiveness of this laterally-constrained stretching mechanism is supported by the numerical model, as will be shown in Section 4.4. The deep cyclones in turn remain to contribute to necking down the upper LC, and this can lead directly to LCE separation.

In contrast, north of 25°N and away from the steep Campeche and Florida slopes, meanders can develop all around the unconstrained periphery of the LC or of a large detached LC ring (Donohue et al., 2016b; Hamilton et al., 2016, 2019). Comprehensive studies on the eddy shedding events Ekman, Franklin, Hadal (Donohue et al., 2016b), and Thor (Johnson et al., 2022) have demonstrated the primary role of baroclinic instability. From our present study of the less-documented DSC region, we provide evidence that both baroclinic joint development and the constraints of side-walls can facilitate LCE separation in the DSC.

Within the DSC, the LC may shift slightly to the Campeche or Florida side as it advances, which restricts meandering along that respective flank. While the deep topography (1500 − 3000 m) is nearly vertical on both sides, the upper topographic slope (100 − 1400 m) is gentle on the Campeche Bank side (Figure 1). The bottom slope there ( ∼120) nevertheless exerts a strong constraint, and the flow closely follows the topography in three case studies presented wherein the LC leans left (west, e.g. Figure 4A). In such instances, LC meandering and deep eddy formation then occurs along the opposite side (toward the Florida Shelf). Case studies presented below indicate that a deep cyclone may be spun up as the LC front meanders away from either of the DSC walls. In the next Section, of the four cases during which eddies were shed, two are strongly influenced by this paradigm. In the other two cases the LC protrudes far enough northward beyond the constraining side walls of the DSC to behave in a manner indicative of baroclinic joint development.

Here we highlight four time periods from the model output showing the evolution of the LC as it extends into the GoM and an LCE separates. Each of the following cases is illustrated by a pair of figures:

Our focus is on deep cyclones that develop within the DSC, as they directly contribute to the necking down and eddy shedding process, and the deep cyclones can oppose the reattachment of detached upper eddies. While baroclinic instability is one process known to drive deep cyclones and anticyclones in the open GoM (Donohue et al., 2016b; Oey, 2008), the results here also emphasize LC interaction with the DSC walls.

In this study, four cases are chosen from a relatively high-resolution NCOM 6-year free run to illustrate strong interactions between the advancing and meandering upper LC and deep eddies. These six model years highlight four southern separation events. The deep cyclones are a persistent feature northwest of the Yucatán Strait inflow in the deep southeastern basin between the Campeche Slope and Florida Slope, which we refer to as the Deep Southeast Channel (DSC). Four cases where LC eddy shedding occurs illustrate important interactions between variability in the upper baroclinic jet and deep eddies whose reference currents are nearly depth-independent throughout the water column. When a LC meander trough recedes from the steep DSC walls, the water column is stretched by the sloped LC pycnocline. The DSC side-walls (the Campeche Slope to the west and the Florida Slope to the east) constrain the deep eddies from translating laterally, allowing these deep features to exhibit a persisting influence upon the upper LC. Indeed, RAFOS and APEX profiling float observations have reported deep mean cyclonic circulation in the DSC (Pérez-Brunius et al., 2018), and the model results demonstrate that a deep mesoscale eddy with nearly depth-independent structure below ∼1200 m spans most of the channel width. The relatively few observations available from the DSC show it is a favorable location for deep cyclones.

The importance of LC interaction with topography has long been recognized, with particular attention to the northern Campeche Bank, the Mississippi Fan, and the West Florida Shelf (e.g. Hurlburt and Thompson, 1982; Chérubin et al., 2006; Le Hénaff et al., 2014; Nickerson et al., 2022; Ge et al., 2025). As the LC flows northward off the Campeche Bank, small meanders form and deepen, coupled with deep perturbations that intensify through vortex stretching (Hurlburt and Thompson, 1982; Chérubin et al., 2006). At the Mississippi Fan, squashing and re-stretching processes occur, again coupling deep eddies with upper-ocean meanders (Le Hénaff et al., 2014). Liu et al. (2016) identified a “pressure point” along the West Florida Shelf near the Dry Tortugas, where contact between the LC and the shallow shelf acts to anchor the current. As long as the LC remains in contact with the shelf at this location, it cannot progress further into the Gulf. Olvera-Prado et al. (2023) also focus on the importance of upper-deep coupling in the steep regions of eastern GOM, particularly the West Florida Slope. Yang et al. (2020) use a multi-scale eddy-mean energetics analysis to demonstrate upper-deep joint development in the deep eastern Gulf where they show a baroclinic energy transfer from mean to eddy fields. This current paper presents another instance of LC–topography interaction in a region that has been largely overlooked, yet it appears to play a significant role in the eddy shedding cycle.

In this free-running realistic model the deep cyclones develop frequently, dominate the local circulation, persist for several weeks, and can facilitate the separation of LCEs. Therefore, we suggest the DSC as a suitable location for near real time observations of the deep flow, using Lagrangian and moored methods, to be assimilated into forecasting models. The central pressure low of the deep cyclone is difficult to detect via satellite and upper-ocean observations, but APEX floats tend to recirculate within this region, and some remain for months. The water velocity at the ∼1500 m float parking depth can be estimated from lateral offsets between successive vertical profiles, and the time-varying deep field could be mapped from a suitable set of floats. Additionally, because the location of cyclone development appears to be so consistent, deep current moorings, bottom pressure sensors, or CPIES, could be placed at a small number of sites to help map the deep flow, and their data could be harvested and relayed ashore by satellite.