Turbulent flow reduces oxygen consumption in the labriform swimming shiner perch, Cymatogaster aggregata

The complex habitats in which many marine organisms live are governed by stochastic, multiscale processes. However, unpredictability is often purposefully limited in studies of fish–flow interaction. Water tunnels, flumes and other apparatus are usually fitted with honeycombs or grids that straighten flow streamlines and minimize turbulent velocity fluctuation (Bainbridge, 1958; Bell and Terhune, 1970; Webb, 1975; Steffensen et al., 1984; Hove et al., 2000; Drucker and Lauder, 2002).

In the interest of mimicking the natural environment, recent laboratory experiments have sought to explore the relationship between fish and more complex flows (as reviewed in Liao, 2007). Flows have been altered via the introduction of boulders (Shuler et al., 1994) and logs (McMahon and Hartman, 1989); Plexiglas structures (Fausch, 1993); cones, spheres and half-spheres (Sutterlin and Waddy, 1975); horizontally or vertically oriented circular cylinders (Sutterlin and Waddy, 1975; Webb, 1998; Montgomery et al., 2003; Cook and Coughlin, 2010; Tritico and Cotel, 2010); and D-section cylinders (Liao, 2003, 2004, 2006; Taguchi and Liao, 2011). Other researchers have introduced fluctuations in water flow speed (Enders et al., 2003; Roche et al., 2014) or body-scale streamwise vortices (Maia et al., 2015). All these methods have served to introduce regular hydrodynamic perturbations in otherwise straightened flows, providing a more realistic approximation of the natural habitat of the animals studied. In some specific cases, the consistent flow features produced by these perturbations are exploited by fish to reduce the metabolic cost of swimming (Enders et al., 2003; Liao, 2004).

The response of fish to these ‘altered’ or complex flows (i.e. non-turbulent but vortex perturbed) can be highly variable (e.g. Cotel and Webb, 2012). It has been hypothesized that swimming in unsteady flows increases energy consumption by requiring additional swimming manoeuvres (Blake, 1979; Weatherley et al., 1982; Puckett and Dill, 1984; Webb, 1991; Boisclair and Tang, 1993; Maia et al., 2015). While some studies have shown that fluctuating flows reduce maximum swimming speed (Pavlov et al., 2000; Tritico and Cotel, 2010; Roche et al., 2014) and increase energy expenditure (Enders et al., 2003; Maia et al., 2015), other studies have found no effect of flow variability on performance metrics in a variety of species (Ogilvy and DuBois, 1981; Nikora et al., 2003). The effects of complex flows on fish swimming performance and oxygen consumption rate are dependent on the magnitude of the turbulence/velocity fluctuations (Pavlov et al., 2000; Lupandin, 2005; Tritico and Cotel, 2010; Webb and Cotel, 2010; Roche et al., 2014) and are probably tied to the fish's swimming behaviour [body–caudal fin (BCF) versus median–paired fin (MPF)]. Swimming performance may also be improved if fish can sense and respond to periodic vortex structures (Liao, 2007). However, these controlled situations may not be directly comparable to many natural habitats in which fish live, where aperiodic components of the flow can be dominant.

It is important to note that many of these ‘altered’ flows do not fit the classical definition of turbulence; they include strong periodic components and little stochasticity, which is a key feature of turbulence (Tennekes and Lumley, 1972; Pope, 2000). In the biological literature, turbulence is described as ‘highly irregular’ (Vogel, 1994) and ‘chaotic’ (Denny, 1993). Therefore, turbulence should be distinguished from highly periodic coherent structures such as waves or vortex streets. These structures may be influenced by turbulence or have turbulence superimposed upon them, but are not themselves ‘turbulent’.

How fish respond to unstraightened, irregular flows in a swimming respirometer, as compared with the straightened and fluctuation-minimizing flows often used in laboratory experiments, is largely unknown. In most previous studies involving ‘altered’ flows (above), periodic fluctuations were introduced by adding features to otherwise straightened flows. In the present study, an unstraightened flow was compared with a standard, straightened flow. In both flows, energy was introduced by an impeller, and the resulting turbulence was allowed to decay over the length of the respirometer. In the unstraightened (high-turbulence) condition, flow was unimpeded from the impeller to the working section, save for a large-opening grid that served to confine the animal to the working section of the swim tunnel (the section to which the fish is confined when swimming; Ellerby and Herskin, 2013). Animals in the working section therefore encountered a field of aperiodic vortices and eddies which remained large relative to the size of the chamber. In the straightened (low-turbulence) condition, the flow features were quantifiably reduced in size via the inclusion of a honeycomb flow straightener and subsequent fine-mesh grid. The flow straightener included tubes with a sufficiently small diameter to damp out velocity fluctuations and eliminate large eddies before they are advected into the test section (Seo, 2013). In the unstraightened flow, turbulence intensity, turbulent kinetic energy, dissipation rate and other metrics were substantially increased relative to the straightened flow. These two configurations produced two distinct conditions: high-turbulence flow (HTF) and low-turbulence flow (LTF).

The metabolic cost of swimming in these two flows was examined, and the swimming kinematics were compared between the two conditions. Although previous studies have shown variable effects of turbulence on the metabolic cost of swimming (e.g. Pavlov et al., 2000; Lupandin, 2005; Tritico and Cotel, 2010; Webb and Cotel, 2010; Roche et al., 2014), it was expected that oxygen consumption rates would be greater in HTF compared to LTF, because of additional postural control and unsteady motion necessary to respond to aperiodic fluctuations in the flow field.

PIV was used to investigate and quantify turbulence in the swim-tunnel respirometer (Fig. 1). The vertical midplane of the working area was illuminated by a 1 W laser of wavelength 445 nm, spread into a thin sheet via a rigidly attached cylindrical lens (S3 Spyder III Arctic, Wicked Lasers, Shanghai, China; Fig. 1). At speeds >0.30 m s⁻¹, images of the vertical midplane were collected at 872 Hz with an exposure time of 1146 μs, using a high-speed, high-resolution camera (HHC Mega Speed PRO X4, Mega Speed, San Jose, CA, USA). At speeds <0.30 m s⁻¹, images were collected at 500 Hz with an exposure time of 2000 μs. Image resolution was 1280×720 pixels at higher speeds and 1280×1024 pixels for lower speeds.

The flow was seeded with hydrated Artemia cysts (Sanders' Premium Great Salt Lake Artemia Cysts) at an approximate seeding density of 16 cm⁻² throughout the illuminated midplane. To ensure neutral buoyancy of the tracer particles, dry Artemia cysts were mixed into beakers of seawater at approximately 40 g l⁻¹, and left undisturbed for a minimum of 40 min to allow positively or negatively buoyant particles to rise or fall out of suspension (Lauder and Clark, 1984). Positively buoyant particles were skimmed from the water surface, and suspension containing near-neutrally buoyant particles was removed and added to the swim-tunnel respirometer. Artemia cysts from the Great Salt Lake strain are known to have a mean diameter of approximately 250 μm (Vetriselvan and Munuswamy, 2011), which in the camera view corresponds to approximately 1 pixel. Light scattering from these tracers increased the average diameter seen in the camera view to 2–3 pixels, as is appropriate for PIV (Melling, 1997; Raffel et al., 2007).

Images were batch processed in Adobe Photoshop to adjust balance and enhance contrast before performing vector computation via two-pass iteration in DaVis (LaVision GmbH, Goettingen, Germany). Subwindows were 64×64 pixels (13.4×13.4 mm) and 32×32 pixels (6.7×6.7 mm) with 50% overlap, resulting in flow resolved to 3.3 mm. This is sufficient to examine the scales of interest, as flow structures smaller than 3.3 mm are not likely to affect fish of the size used in this experiment (Tritico and Cotel, 2010). Vectors were computed for the entire working area in both HTF and LTF.

Several of the many parameters available to describe the level of variability in turbulent flows (Cotel and Webb, 2012) were calculated for both HTF and LTF. These parameters are defined and described in detail in Table S1. They are u_T, the turbulent velocity scale; TKE, the turbulent kinetic energy; TI, the turbulence intensity; ω_y², the enstrophy (the square of vorticity); ε, the energy dissipation rate; L_xx and L_zz, the integral length scales in x and z; λ, the Taylor microscale; and η, the Kolmogorov microscale. All parameters (with the exception of the integral length scales and spectra) were calculated as spatiotemporally varying quantities, defined at each x–z point for each frame (where a frame is one PIV image pair). Quantities were averaged in time to calculate mean quantities, with 95% confidence intervals calculated via bootstrapping in time (approximately 6 s per flow speed and turbulence condition, with 8 data points calculated throughout the range of tested speeds). The outer 2 cm of each frame (close to the grid and walls) was not included in the averages to avoid including the effects of boundary layers. Integral length scales were found by calculating the appropriate autocorrelations in x and z. Throughout the manuscript, we use u to represent the spatiotemporally varying fluid velocity, whereas U represents the swimming speed of the fish (equivalent to the mean fluid velocity across the tunnel).

Non-overlapping 95% confidence intervals (CIs) indicated that HTF had significantly higher average values of TKE, TI, ω_y² and ε compared with LTF (Fig. S1), across the range of tested flow speeds. The difference between HTF and LTF generally increased with flow speed: TKE, ω_y² and ε all increased with flow speed in HTF, but in LTF, these parameters remained low and relatively constant despite the increasing flow speed. TI was higher in HTF than in LTF but was highest at low speeds and reached a plateau as speed increased (0–0.4 m s⁻¹; Fig. S1D). This was expected, as TI represents the ratio of fluctuations to mean flow; as mean flow increases, fluctuations will become stronger and TI will remain constant.

At high speeds, TI was approximately 150% higher in HTF than in LTF. At the highest tested speed, TKE was 500% higher in HTF, ω_y² was 170% higher and ε was 370% higher. For the difference in turbulence properties across the full range of flow speeds tested, see Fig. S1. The turbulence properties observed in the laboratory-generated flow (in both HTF and LTF) were comparable to those that would be experienced by C. aggregata in its natural habitat of coastal estuaries, bays and streams (see Table 1).

Three turbulent length scales were calculated to illustrate the differences in eddy size between HTF and LTF. The Taylor microscale λ (representing the average distance between stagnation points within the flow) and streamwise integral length scale L_xx (representing the overall average eddy size) were both larger in HTF than in LTF, and the Kolmogorov scale η was generally smaller in HTF (Fig. S2). [We note that the Kolmogorov scale in our experiments, averaging approximately 0.2–0.4 mm, is aligned with the expected values for small-scale ocean turbulence (0.3–2 mm; Jiménez, 1997). However, it is not possible to match the larger length scales to the animal's natural habitat. In an experimental context, these larger length scales are bounded by the respirometer's overall size and therefore cannot approach the wind- and tide-driven scales typical of the coastal ocean.] The relative sizes of these three scales signify that in HTF, fish were swimming through a larger range of spatial scales within the flow. Additionally, the vorticity field suggests differences in overall flow structure between the two flow conditions (Fig. 2). To verify that the flow did not contain any significant periodic components, the frequency spectra of both HTF and LTF were calculated throughout the tested velocity range (Fig. S3). No discrete periodicity, such as that produced by vortex shedding or driven by the flow impeller, was observed within the working section, and the spectra were consistent with classic turbulent spectra showing an energetic cascade from low to high frequencies. As expected, HTF contained more energy overall across all frequencies (Fig. S3).

To summarize, two different flow conditions were created in a swimming respirometer. Both of these flows are turbulent, containing aperiodic velocity fluctuations. Neither flow contains large, periodic coherent structures (such as the von Kármán vortex street typically shed from a bluff body; Fig. S3). However, HTF is measurably more turbulent than LTF: it has larger velocity fluctuations at a given speed (higher TI), dissipates energy at a higher rate (ε) and contains a larger variance of eddy sizes (L_xx, L_zz, λ, η) as well as stronger extremes of vorticity (ω_y²; Figs S1 and S2). Therefore, a significant difference was expected in the energetics of fish swimming in these two flow conditions.

Shiner perch, Cymatogaster aggregata Gibbons 1854, were collected by beach seine at Jackson Beach, San Juan Island, WA, USA (48°31′N, 123°01′W) in July and August 2013. Fish body length (BL, mean±s.d.) measured 12.1±0.3 cm, and mass was 40±7.1 g. Fish were kept in 120×56×15 cm and 90×59×30 cm tanks with recirculating seawater at the University of Washington's Friday Harbor Laboratories, and they fed on plankton present in the water. Fish were fasted for a minimum of 4 h prior to experimental trials. Ambient seawater temperature followed ocean conditions of the area and ranged from 12 to 14°C. After capture, individuals were maintained in aquaria for at least 3 days before their first experimental trial, and were given a minimum of 2 days to recover between trials. Eight fish that performed trials in both the low- and high-turbulence conditions were used for analysis. The order of low- and high-turbulence exposure was randomized for each individual by flipping a coin.

Experiments were conducted in an 8.31 l clear Steffensen Model 1 Plexiglas swimming respirometer with a working section of 9.0×26.0×11.0 cm (width×length×depth, w×l×d). LTF and HTF were induced by inserting two different grid conditions (Fig. 1). For LTF, the tunnel was fitted with a 9.4×10.6×3.3 cm (width×depth×thickness, w×d×t) honeycomb straightener (0.6 cm tube diameter) followed by a 0.01 cm² square mesh mounted on a 9.2×11.9×1.7 cm (w×d×t) plastic grid (1.2×1.2 cm opening size). This fine mesh reduced the size of the incoming coherent flow structures, further straightening the flow. For HTF, the tunnel was fitted with a 9.2×11.8 cm (w×d) large-opening grid [opening size 2.45×2.65 cm (l×h), coated wire thickness 2.7 mm on average]. These large openings constrained the fish to the desired test section while still allowing relatively large eddies (approximately 60% of fish body depth) to enter the test section. The honeycomb straightener was not present in the HTF regime.

Flow speed within the working section was driven by an impeller attached to an AC motor and motor control (DRS71S4/FI and Movitrac LTE-B B0004 101-1-20, respectively; SEW Eurodrive, Wellford, SC, USA), and was calibrated prior to experiments with a Höntzsch TAD W30 flow-meter (Höntzsch, Waiblingen, Germany). Flow speed is hereafter reported in body lengths per second (BL s⁻¹). Water temperature during trials was maintained between 12.9 and 13.0°C by an external chiller.

Before each trial, fish were acclimated in the test section for a minimum of 4 h (maximum 8 h 53 min), swimming at 0.5 BL s⁻¹. After acclimation, speed was increased by 0.5 BL s⁻¹ every 30 min up to a maximum of 4.5 BL s⁻¹ (the maximum speed reached by any fish; maximum of 9 speed measurements in total for a given fish). Solid blocking effects were accounted for based on criteria established by Bell and Terhune (1970). Each fish performed LTF and HTF trials in a randomized order; the time of day of the two trials was held constant for a given fish to eliminate potential effects of photoperiod on fish metabolism and allow direct comparison between HTF and LTF for the same individual. The rate of oxygen consumption (Ṁ_O2; mg O₂ kg⁻¹ h⁻¹) was obtained through intermittent flow respirometry (Steffensen, 1989). Oxygen was measured with a PreSens Fibox 3 fibre optic oxygen meter (PreSens, Regensburg, Germany) inserted into the flushing chimney, and Ṁ_O2 was calculated in LoliRESP as described in Steffensen (1989). Three oxygen determinations were made at each swimming speed, with 180 s flush, 120 s wait and 600 s measurement periods comprising a 15 min cycle (Svendsen et al., 2016). Following each trial, oxygen consumption was measured in the closed, empty respirometer to obtain background bacterial respiration rates. The measured bacterial respiration was then subtracted from all Ṁ_O2 measurements in that trial. Bacterial respiration was not measured for two trials (fish 5 LTF, fish 13 HTF); in those cases, the average bacterial respiration from all other trials was subtracted from total measured Ṁ_O2.

All swimming kinematics, including both fin beat frequencies and fish position within the working section, were filmed at 30 frames s⁻¹ with a tripod-mounted Casio Exilim EX-FH100 camera filming horizontally. A mirror was placed above the respirometer's working section at 45 deg to provide simultaneous lateral and aerial views of the swimming fish, similar to the schematic diagram shown in Fig. 1. Kinematics were measured at three distinct swimming speeds: low (0.5 BL s⁻¹), medium (1.5 BL s⁻¹) and high (shared maximum velocity, U_max; range 3.5–4.5 BL s⁻¹, mean 3.9 BL s⁻¹). High swimming speed is defined as the shared U_max, being the highest flow velocity attained by a given individual before fatigue in both low- and high-turbulence trials (e.g. if a fish attained 3.5 BL s⁻¹ in LTF and 4.0 BL s⁻¹ in HTF, the shared U_max is 3.5 BL s⁻¹). Fatigue was defined as when a fish could no longer swim unassisted and rested against the back grid for >5 s; a speed level was ‘attained’ if a fish performed all flushing cycles at that flow speed. These three levels were chosen to illustrate swimming kinematics throughout a range of speeds and to simplify analyses relative to the more highly resolved Ṁ_O2 analysis.

The 3D position coordinates of the tip of the snout of each fish were digitized in Tracker v.4.81 (http://physlets.org/tracker/). Coordinates were digitized every 10 frames (equivalent to 3 frames s⁻¹) for the last 180 s of the second (of three) 600 s measurement periods. For each trial, the centroid (x_c, y_c, z_c) and standard deviation of the 3D snout coordinate (x_s, y_s, z_s) were calculated for each trial, with the standard deviation reflecting some measure of variance overall (e.g. Fig. 3).

Pectoral and caudal fin use were determined from video footage at the previously defined low, medium and high flow velocities. The time of adduction of the pectoral fin and of complete cycles of left and right displacement of the caudal fin was recorded (e.g. Drucker and Jensen, 1996). We calculated the pectoral and caudal fin beat frequencies (Hz) as the reciprocal of the mean waiting time between beats (1/mean difference in beat time) over the 180 s measurement period (Fig. S4). The gait transition speed (U_pc) in C. aggregata (12 cm) has been reported to be ∼3.5 BL s⁻¹ (e.g. Mussi et al., 2002) and we observed some use of the caudal fin, in line with previous work. However, a transition to exclusively caudal fin locomotion was not observed in any fish.

To assess how Ṁ_O2 changed with fin beat frequency, we fitted a generalized additive model (Wood, 2017) to the Ṁ_O2, using pectoral or caudal fin beat frequency (Hz) as separate explanatory variables. We used factor–smooth interactions (Baayen et al., 2018) to fit two levels of a smooth function of measured beat frequency, one for each flow condition. Such terms fit the base level of the factor as a smooth, then model deviations from that smooth for the other level; thus, information was shared between the models while allowing for a flexible relationship that makes no assumptions about functional form. Fish ID was included as a random effect. Models were fitted by restricted maximum likelihood.

Goodness of fit for the generalized additive model of oxygen consumption as a function of pectoral fin beat frequency was assessed using standard plots (Wood, 2017). The factor–smooth interaction was sufficiently flexible to model the shape of the relationship (6.489 effective degrees of freedom, given a maximum basis complexity of 20), deviance residuals appeared to be approximately normal and showed little pattern with increasing values of the linear predictor (hence did not have an issue with heteroscedasticity), and the relationship between fitted and observed values of Ṁ_O2 was approximately linear. The random effect for fish ID approximated normal from a Q–Q plot.

We attempted to fit a similar model to the relationship between oxygen consumption and caudal fin beat frequency; however, the model fit was unsatisfactory as 64.6% of the time the fin beat frequency was recorded as zero, limiting model fitting options. Given the poor fit of our model, we do not present any modelling results for caudal fin beat frequency but present the raw data in Script 1.

Following Kerr et al. (2016), we visualized the 3D position of all fish in each flow and speed condition with heatmaps, to determine whether fish were consistent in their positions in HTF or LTF, perhaps to take advantage of regional flow conditions or ‘dead spots’. To aid interpretation of the results, we plotted the standard deviation of position per fish, flow condition and speed and fitted a rudimentary linear model per fish and flow condition (Script 1).

To determine the relationship between fish swimming kinematics, energetics and ambient flow, the frequency components of the fish's x–y–z position within the respirometer were analysed. This analysis was undertaken to further ensure that any periodic position fluctuations displayed by the fish were not a result of variables related to the surrounding flow (i.e. condition or speed); that is, that periodic position fluctuations were not indicative of fish using vortex structures to save energy. Periodicity in the time series of (x_s, y_s, z_s) (e.g. Fig. 3B–D,F–H) was detected for each fish with Fisher's g-statistic above 0.1 Hz as a result of the 3 frames s⁻¹ sampling frequency (Wichert et al., 2004). Space-use analyses were completed in MATLAB (R2011a-R2014b, MathWorks, Inc., Natick, MA, USA).

At 0.5 BL s⁻¹, fish in the HTF condition consumed 10.5±6.4 mg O₂ kg⁻¹ h⁻¹ (66%) less than when in the LTF condition (167.2±15.6 mg O₂ kg⁻¹ h⁻¹). The exponent of the relationship was significantly different between these flow conditions (t₃₆₉=−2.328, P=0.0204); differences between treatments occurred at speeds >2 BL s⁻¹. Ṁ_O2 was not significantly different between LTF and HTF at speeds below 1.5 BL s⁻¹ (paired t-test with false discovery rate correction, F₇<1.889, P>0.13; Fig. 4). Fish consumed significantly less (on average 20% less, range 0–46%) O₂ in HTF versus LTF conditions at speeds above 2.0 BL s⁻¹ (F₇>3.038, P<0.03; Fig. 4).

As expected (e.g. Roche et al., 2014), predicted Ṁ_O2 monotonically increased with pectoral fin beat frequency (Fig. 5) in both flow conditions. The factor–smooth interaction showed significant divergence between the two flow conditions when the fin beat frequency was above 2.75 Hz, as evidenced in the difference in 95% CIs (±2 standard errors from the smooths of fin beat frequency; Fig. 5).

Overall space use in the tank was visualized using heatmaps, following Kerr et al. (2016), which did not suggest consistent positioning in the tank between fish and in different flow and speed conditions (Fig. 6). The centroid positions in x, y and z showed no consistent response to speed or flow condition (Script 1). In both flow conditions in the x- and y-directions, deviation decreased with increasing speed. In the x- and y-directions, deviation was greater in HTF than in LTF. For the z-direction, there was little difference in deviation between LTF and HTF conditions. Further, there was little effect of speed on deviation in HTF but a stronger effect of speed on deviation in LTF (Script 1). In combination with heatmaps (Fig. 6), this approach gives some idea of the behaviour of the fish in the tank. Overall, space use in LTF was constrained to a small area at the front of the working section. In HTF, fish position was spread across a larger area (Fig. 6).

Significant periodicity was detected in time series of x_s (1.81±2.95 Hz; mean±s.d.), y_s (1.57±0.76 Hz) and z_s (11.35±0.66 Hz), but in only a proportion of all cases (x=26/48, y=34/48, z=32/48).

Most studies of fish swimming and respirometry have used flow-straightening devices that result in low-turbulence flow conditions within the experimental setup, similar to the LTF condition described in this study. The natural habitat of most fishes is more turbulent than that created by such laboratory conditions, but most studies seeking to create more natural flows have focused on fish behaviour in coherent vortex structures, which represent only a small portion of the flows encountered by fish in their natural habitats (Lacey et al., 2012; Kerr et al., 2016). In predictable vortex-dominated flows, it is common for fish to display energy-saving behaviour (e.g. Kármán gaiting) as a result of the consistent nature of the periodic vortices (Liao, 2003; Liao et al., 2003). Few studies have explicitly tested the effect of aperiodic, more randomized flow on fish swimming energetics and kinematics. Of these, many have suggested that fish should expend more energy in unsteady flows (Enders et al., 2003; Lupandin, 2005; Tritico and Cotel, 2010) or show greater variation among individuals (Kerr et al., 2016), as each fish must accommodate the unique flow structures it encounters. In the experiments described herein, we compared oxygen consumption and swimming movements of fish in LTF (similar to a standard fish respirometry study) versus a HTF (mimicking a more natural turbulent environment). Fish displayed significantly reduced metabolic costs when swimming in HTF compared with LTF. No periodic components of the flow or swimming kinematics can account explicitly for these energy savings, suggesting that fish may also be able to exploit turbulent flows without discernible periodic wake formations.

Fish have been shown to reduce the energetic costs of swimming (or otherwise adapt to changes in ambient flow) in specific circumstances by employing behaviours such as (1) gait switching (Korsmeyer et al., 2002), (2) Kármán gaiting (Liao, 2003), (3) entraining (Webb, 1998), (4) bow riding (Newman and Wu, 1975; Taguchi and Liao, 2011), (5) tail holding (Kerr et al., 2016) and (6) wall holding (Kerr et al., 2016). Below, the present results are discussed in the context of expected fish behaviour under these different energy-reducing strategies to identify potential mechanisms behind the reduced metabolic costs measured in this study.

We found that Ṁ_O2 increased significantly more with pectoral fin use in LTF than in HTF, and caudal fin use was often absent in these MPF swimmers. At high speeds prior to exhaustion, many MPF swimmers such as C. aggregata and other labriform swimmers start complementing pectoral fin swimming with the caudal fin (i.e. at gait transition; Webb, 1973; Svendsen et al., 2010). Webb (1973) describes little-to-no caudal fin use in C. aggregata at speeds below 3.4 BL s⁻¹. Typically, at speeds of 3.5–3.85 BL s⁻¹, C. aggregata used occasional low-frequency, low-amplitude caudal fin beats (caudal fin pattern A; Webb, 1973). Above 3.85 BL s⁻¹, C. aggregata used 1–3 caudal fin beats in quick succession to maintain position in the swimming flume (burst–coast swimming; caudal fin pattern B; Webb, 1973), but this occurred over a short period of time, immediately prior to exhaustion. Therefore, the overall proportion of time the caudal fin is used even after initial recruitment of the caudal fin can remain low (Webb, 1973; this study); gait transition in other labriforms has been shown to involve initial occasional recruitment of the caudal fin before slowly developing to full and continuous burst–coast swimming (Cannas et al., 2006). In addition, the fish in our study were smaller (12.1 cm) than those used by Webb (1973; 14.3 cm), and therefore they are expected to start using the tail at a higher relative speed (in BL s⁻¹; Mussi et al., 2002) than that found by Webb (1973). Further experiments focusing on fin and muscle use (pectoral versus caudal fin and red versus white muscle; e.g. Gerry and Ellerby, 2014) in unpredictable flow regimes would better resolve the strength of the relationship between Ṁ_O2 and movement patterns and the mechanism behind different movements in complex flows.

Fish can reduce energetic expenditure by stationing behind a bluff body and altering their body kinematics to synchronize with shed vortices. This behaviour, called Kármán gaiting (Liao, 2003), requires the presence of a bluff body which can shed vortices of a size of the order of the fish's body depth (Tritico and Cotel, 2010); no such bluff body was available in the present experiments. The cross-tank body oscillation displayed by fish in HTF (e.g. Fig. 3) is at first glance suggestive of such a behaviour; however, this behaviour was not consistent between all fish, and the flow did not offer discrete periodicity (Fig. S3).

The orientation and/or size of vortices shed behind moving fins of MPF (sensu Webb, 1984) swimmers differs from that of BCF (sensu Webb, 1984) swimmers (Drucker and Lauder, 1999, 2002; Drucker et al., 2005; Fish and Lauder, 2006). Similarly, MPF swimmers maintain a rigid body during locomotion at speeds below gait transition and therefore would not be able to use energy-saving behaviour in the same manner as BCF swimmers, i.e. by synchronizing their tail beats with the vortex-shedding frequency (Kármán gaiting; Liao, 2003). Little is known about how MPF swimmers interact with vortices and turbulence in order to save energy. Interestingly, previous work has shown that, like BCF swimmers (Marras et al., 2015), MPF swimmers can save energy when swimming in the wake of neighbours in a school (Johansen et al., 2010). It is therefore possible that MPF swimmers exploit paired-fin motion kinematics and timing in order to minimize the energy spent on swimming in turbulent flow. Simultaneous kinematic measurements and PIV may better resolve the fine-scale movement responses of fish to moving flow structures.

Further energy-saving mechanisms often involve station holding either in front (i.e. bow riding; Taguchi and Liao, 2011) or behind (i.e. entrainment; Liao, 2006; Przybilla et al., 2010) bluff bodies, where fish use the resulting high-pressure zones or lift and wake suction forces, respectively, to maintain position. Fish have also been shown to tail hold by resting their tails against screens at the rear of experimental setups (Kerr et al., 2016). These behaviours cannot explain the reduced energy consumption observed in this study as there were no bluff bodies to station in front of, and fish generally stayed at the front of the respirometer, with their tails 5–15 cm from the rear grid (see Fig. 6). Fish can also display wall holding by potentially taking advantage of more stable and reduced flows in the wall boundary layer (Kerr et al., 2016). While overall space use (Fig. 6) shows that fish in this experiment had a slight preference for swimming on the left side of the respirometer, this positioning was dynamic in time. Fish moved back and forth across the respirometer at higher flow velocities and in HTF (Fig. 3). Any spatial bias was therefore not likely to be due to standing flow features; that is, based on this movement pattern and PIV, there is no evidence of ‘dead zones’ of low flow where fish could consistently position themselves. In HTF, fish exhibited more movement across and along the test section (Fig. 6). A comparison of the movements of one fish (Fig. 3) and the heatmap of all fish (Fig. 6) shows individuals differ in their absolute position but that their positioning is consistently more variable in HTF. This could indicate a passive behaviour, i.e. fish are being advected back and forth by the larger turbulent eddies found in HTF, or an active behaviour, i.e. fish are exploiting temporary (not periodic) vorticity, which is higher in HTF than in LTF, to their advantage, in line with the lower Ṁ_O2 in HTF versus LTF conditions.

The results presented herein are not explained by previously described energy-saving mechanisms observed in fish. The fact that fish consume less oxygen when in more turbulent conditions may be the result of greater variability in their positions (compared with LTF), which in turn may help the fish to take advantage of the variability in turbulent energy through time and space observed in HTF. However, the mechanism by which this advantage is gained is not clear. The difference in energy consumption may also be due to the interplay between skin friction drag and pressure drag on a given fish. The Reynolds number of each fish ranged from approximately 5×10⁵ to 5×10⁶. This range is close to the ‘drag crisis’ regime, which is well studied in spheres and cylinders (Smith et al., 1999; Singh and Mittal, 2005; Kundu et al., 2011). In this regime, the boundary layer over the surface of a bluff body transitions from a laminar to a fully turbulent state, resulting in a smaller wake and lower pressure drag. In this case, the velocity fluctuations in HTF may ‘trip’ the boundary layer into the turbulent state, lowering the pressure drag compared with LTF. Further investigation of this possibility would require simultaneous PIV of the fish and flow.

The PIV analysis described here characterizes the flow and its eddies with no fish in the tank; the same analysis with the fish swimming simultaneously in the flume (e.g. Drucker and Lauder, 2002) was not possible because of (1) potential alteration of fish behaviour due to the laser light, (2) optical inaccessibility due to the presence of the animal and (3) animal welfare concerns over particulate density and non-infrared laser light. We were therefore unable to examine the specific flow structure for each individual and instead measured a ‘representative’ flow field in each condition (throughout the range of tested speeds) to be used for inference of all trials. Because of this, we cannot exclude the possibility that the presence of the fish created unique flow features, such as persistent zones of lower than average flow. However, if this were true, we would expect to see less-variable positioning in HTF as fish held station in these ‘self-generated dead zones’ to save energy. In fact, we found the opposite: less-variable position (and higher energy consumption) in LTF. Further understanding of the mechanism behind the observed behaviour could be achieved by conducting a similar experiment in which the fish kinematics, the flow field surrounding the fish and the fish's Ṁ_O2 were measured simultaneously.

Both turbulence and coherent vortex structures can have passive and active effects on fish, playing a role in postural control (Pavlov et al., 2000; Tritico and Cotel, 2010), foraging (MacKenzie and Kiørboe, 1995), transportation costs (Webb and Cotel, 2010; Webb et al., 2010), and orientation and swimming speed (Standen et al., 2004; Lupandin, 2005). It is critical that studies continue current research trends to determine the energetic, kinematic and behavioural effects of swimming in non-uniform, aperiodic flows that mimic the diversity of turbulence observed in habitats. In particular, this study has found that fish show significantly different patterns of positioning, kinematics and energetics in a typical laboratory flume (LTF) versus a more turbulent flow (HTF), at the highest flow velocities. Individual and context-specific responses to variable flows in terms of propulsive (Liao, 2003, 2004, 2007) or positioning strategies (herein) must be understood to better relate laboratory-based findings to natural environments (Roche et al., 2014). Whether the effects of turbulence on energy consumption are positive, neutral or negative (Enders et al., 2003), their quantification is essential in understanding the energetics of swimming in semi-natural, varying flow conditions (Cotel and Webb, 2015).

We would like to thank UW Friday Harbor Laboratories for their facilities, space and support for this project, Mega Speed for the loan of the high-speed camera, and E. A. Variano (UC Berkeley) for the loan of PIV laser equipment.