Introduction

In a world with a perpetual need for freshwater but a finite capacity for greenhouse gas emissions, reducing the energy consumption associated with water purification is imperative. While the energy requirements of water desalination have decreased considerably in recent decades due to improvements in processes and membranes1, several challenges remain, including brine management, membrane fouling, and the high osmotic pressure of the most prevalent source water (seawater). Desalination and reuse processes that can treat low-salinity wastewater at high recovery (and without membrane fouling) have the potential to produce clean water at much lower energy consumption and cost than seawater desalination2. Dynamic processes such as batch and semibatch RO have emerged as an approach to reducing energy consumption and increasing water recovery3,4 while avoiding inorganic fouling (scaling)5,6,7.

A true batch RO system, into which no feed enters during permeate production, has the potential to have the lowest energy consumption and scaling risk of any RO variant3,8. Membrane scaling, or the formation of inorganic solids on the membrane surface, requires a locally supersaturated solution to be present for a sufficiently long time for crystal nuclei to form9. In conventional RO, scaling risk is highest at the end element where membranes are constantly exposed to the highest salt concentration10, allowing ample time for scalants to nucleate. Several dynamic RO processes (including the most mature dynamic process, semibatch RO) have the potential to combat scaling by regularly exposing membranes to subsaturated conditions between cycles and limiting the duration of supersaturated conditions. However, semibatch RO has higher average retentate concentration than batch RO because it constantly adds saline feed to the system during permeate production, thus linearly increasing salinity11. Spending more time at lower concentrations in batch RO means lower average osmotic pressure and thus the potential for lower energy consumption; it also means less time at the most supersaturated conditions that are most prone to scaling. The recovery of a dynamic RO process such as batch RO can be tuned so that the membrane is exposed to supersaturated solution for a shorter period than the nucleation induction time—the time delay before crystal nuclei can be observed12—of the most likely scalant6. Several batch RO designs13 have been explored to realize these potential advantages, including configurations utilizing a bladder7,8, piston11,14, or counterflow pressure exchanger15,16.

This study investigates the use of a pilot-scale batch RO prototype with a flexible bladder to recover water from the concentrate produced by an existing RO facility. High-recovery desalination has been tested at the laboratory scale with batch RO (e.g., refs. 8,17) and at the pilot scale with semibatch (also known as closed-circuit) RO18. However, to the authors’ knowledge, there are not yet any published pilot-scale studies of true batch RO using real source water. In this study, a bladder-based batch RO system was designed to enable energy-efficient, high-recovery batch operation by allowing high-pressure circulation of a shrinking volume of retentate. The batch RO system was then tested on site at a facility desalinating agricultural drainage water, where it recovered water from the facility’s concentrate stream. The pilot system’s recovery ratio, permeate quality, and energy consumption are analyzed to evaluate the performance of bladder-batch RO in desalination of real RO brine.

Results

During a week of nearly-continuous operation at the Yuma Desalting Plant (YDP), the batch RO pilot system completed 885 batch cycles and produced 31.1 m3 of permeate with a mean salinity of 150 mg/kg total dissolved solids. This section describes the behavior of the pilot system, its water quality, energy, and scaling performance, and qualitative observations of system operation.

Figure 1 shows the behavior of the pilot system during a single representative batch cycle. During each cycle, permeate was produced for about 9 min, as shown by the flux curve, which rises quickly and then remains level at about 15 L/m2 h (LMH), equal to the high-pressure pump’s flow rate. Permeate production was followed by about half a minute of flushing out brine and 1 min of recharging the system with fresh feed; the duration of each phase was determined by the control system described in section “System control”. Feed was introduced to the system during flushing and recharge only, as shown in Fig. 1b. During permeate production, the feed pressure was not sufficient to introduce any feed to the system; however, the feed flowmeter’s voltage tended to drift, so the feed flowmeter reading was offset by a fixed value in in each of Figs. 1b and 2b to correct it to zero during the permeate production phase.

Fig. 1: Operation data for a typical batch RO cycle.
Fig. 1: Operation data for a typical batch RO cycle.
Full size image

a Pressure and (b) flow data during a single cycle of batch RO system operation with a feed of YDP concentrate. “Pressure drop” refers to the pressure drop in the retentate between ends of the bladder pressure vessel. Vertical dashed lines show transitions between phases of operation.

During permeate production, the pressure rises rapidly (over about 15 s) until flux reaches its desired level, dips slightly (likely as remaining brine is pushed from the feed channel into the space surrounding the bladder), and then slowly rises throughout the rest of the permeate production phase. The pressure rises more quickly at the end of the permeate production phase as the volume of retentate is minimized, corresponding to faster increase in retentate osmotic pressure. Pressure drops almost instantaneously when the brine valve is opened at the end of permeate production. The same rising, concave-up pressure profile was seen in our group’s previous testing of the pilot system with NaCl solutions7.

Differential pressure across the bladder pressure vessel was used as an indicator of bladder fullness. Differential pressure rises somewhat partway through the cycle (possibly as the bladder begins to touch the tubes installed just inside the pressure vessel’s side ports) and then begins to rise rapidly as the bladder becomes full, triggering the control system to switch to the flush phase. Differential pressure remains high during flush because the bladder is full and then drops quickly at the beginning of the recharge phase while the bladder is allowed to empty.

Figure 1 also shows how the direction of water flux through the membranes reversed during the flush and recharge phases. The negative flux had an initial magnitude similar to the permeate production phase; however, the magnitude dropped quickly as flush and recharge continued. This osmotic backwash occurs because the osmotic pressure in the feed channel exceeds the pressure of the feed entering the system, causing osmosis through the membranes. The rate of backwash decreased as the osmotic pressure in the feed channel decreased, and it may have also decreased due to salt permeation into the permeate channel. This reverse flux of permeate reduced overall recovery, but this was compensated for during pilot testing by setting the permeate production phase’s recovery higher than the desired overall recovery.

The flow and pressure profiles of the pilot system were largely consistent from cycle to cycle. As an example, Fig. 2 shows 3 h of operation that are fairly representative of the behavior observed during the week-long pilot.

Fig. 2: Typical system operation.
Fig. 2: Typical system operation.
Full size image

a Pressure and (b) flow data for 3 h and 17 cycles.

The system’s overall recovery of permeate from YDP concentrate during the week of testing was 82.6%. Recovery was determined based on feed (i.e., YDP concentrate) and batch RO concentrate volumes, which were calculated from Water Quality Improvement Center (WQIC)-provided flow rate sensors integrated over time throughout the week-long pilot period. The ultrasonic sensors measured flow rate independent of direction, so some permeate back-flow may have been logged as positive permeate flow; feed and concentrate only flowed one direction, so those sensor readings were regarded as more accurate. A concentration-based recovery Eq. (1), which is derived from salt and water mass conservation, assuming constant mixture density), is calculated from feed, concentrate, and permeate total dissolved solids (TDS; see Table 1) to be 82.8%, in good agreement with the volume-based recovery. We regard the volume-based recovery of 82.6% as more accurate because flow rate was measured continuously while the concentration-based recovery relies on only two concentrate samples. The recovery was tuned slightly throughout the week between approximately 82% and 84%, as shown in Fig. 3, which shows the recovery during normal operation (i.e., excluding shutdowns) based on the feed consumed and concentrate produced during each 2-h period.

$${\rm{RR}}=\frac{{C}_{c}-{C}_{f}}{{C}_{c}-{C}_{p}}$$
(1)
Table 1 Mean composition of the batch RO system’s feed (i.e., YDP concentrate), concentrate, and product water
Full size table
Fig. 3: Recovery ratio, calculated for each 2-h period from feed and concentrate flow rates.
Fig. 3: Recovery ratio, calculated for each 2-h period from feed and concentrate flow rates.
Full size image

Variations in recovery ratio are primarily due to changes in setpoint.

The overall recovery achieved by the batch RO pilot in combination with the YDP’s primary desalination process is estimated here by incorporating the YDP’s historical water composition data. Recent representative concentrations have been 2217 mg/L in the YDP’s feed and 7650 mg/L in its concentrate, as reported in the More Water Less Concentrate Technical Guidelines, www.morewaterlessconcentrate.org/technical-guidelines.html, accessed Feb. 22, 2022. The YDP’s permeate concentration (which has a minor effect on calculated recovery) was reported to be 252 mg/L in a 2007 demonstration19. Using Eq. (1), the YDP’s recovery is estimated to be RRYDP = 0.734, or 73.4%. Given the use of YDP concentrate as the batch RO system’s feed, the overall recovery RRcombined achieved by the YDP and the batch RO system (with its recovery ratio of RRbatch = 0.826) was approximated using Eq. (2) to be 95.4%.

$${{\rm{RR}}}_{combined}=1-(1-{{\rm{RR}}}_{YDP})(1-{{\rm{RR}}}_{batch})=0.954$$
(2)

The quality of the permeate produced by the batch RO system improved over the course of the pilot period as shown in Fig. 4. During the first 4 days of operation, the permeate TDS decreased. This behavior is consistent with a DuPont RO membrane manual20, which notes it is common for an RO membrane’s rejection to improve over the first several hours to days. The mean TDS of all permeate samples was 150 mg/L. Despite taking the YDP’s concentrate as its feed, the batch RO system produced lower-salinity permeate than that produced by the YDP’s primary desalination process in a 2007 demonstration19.

Fig. 4
Fig. 4
Full size image

Permeate salinity over time, measured by the WQIC as total dissolved solids (TDS).

Energy consumption

Based on the total energy consumption and permeate volume (calculated conservatively from the difference between feed and concentrate volumes), the prototype had an average specific energy consumption (SEC) of 3.3 kWh/m3 during the testing period. SEC and system power fluctuated little throughout the duration of the pilot test except during unplanned outages, as shown in Fig. 5. Excluding the last day, when a laptop crash led to downtime, the average power consumption was 633 W. The pilot’s SEC of 3.3 kWh/m3 is comparable to modern seawater RO plants21, which desalinate higher salinity water but have the efficiency benefits of a much larger scale. However, because the bladder-batch RO process only discharges brine at atmospheric pressure, the small pilot-scale prototype achieved a relatively low energy consumption, even without an energy recovery device.

Fig. 5: Power and specific energy consumption (SEC) reported as daily averages.
Fig. 5: Power and specific energy consumption (SEC) reported as daily averages.
Full size image

Both were affected by an unplanned outage on Day 7.

System pressure and energy consumption were also modeled as described in section “Modeling” to enable comparison with the pilot data and extrapolation to hypothetical larger-scale batch RO systems. The applied pressure increases over time as the osmotic pressure of the retentate rises throughout each batch cycle. The modeled pressure profile was compared to pressure data during a representative batch cycle in Fig. 6, showing reasonable agreement but a slight under-prediction of pressure midway through the permeate production stage. The model assumes the batch RO’s retentate is perfectly mixed when outside the membrane module. However, preferential flow paths around the bladder likely allow some higher-concentration retentate, which has just left the membrane module, to return to the membrane module without fully mixing with the rest of the retentate around the bladder. This higher-concentration retentate entering the feed channel necessitates slightly higher feed pressure to maintain constant flux. Therefore, the higher actual pressure observed is likely due to imperfect mixing in the space around the bladder in this prototype.

Fig. 6
Fig. 6
Full size image

Comparison of modeled and measured hydraulic pressure.

An optimized, large-scale batch RO system has the potential for a much lower SEC by using larger, more efficient pumps. We estimated the SEC of a 100,000 gpd (379 m3/d) batch RO system treating YDP concentrate at the pilot’s recovery (82.6%) using the approach described in section “Modeling”. When pilot performance data was compared to a model of the pilot system (as described above), the average pressure was approximately 8% higher than modeled. Assuming the same mixing issue might persist in a large-scale system, the Swaminathan et al.15 model was modified by adding 8% to the high-pressure pump’s energy consumption when estimating energy consumption at large scale.

A large-scale system would differ from the pilot by using more efficient pumps and having fewer additional energy-consuming components, such as valves, per unit production capacity. Comparing the system performance model to the actual power consumption of the two pumps during testing, the high-pressure pump was operating at an efficiency of approximately 40% and the circulation pump was operating at an efficiency of only about 1%. The low efficiency is due to the circulation pump’s use at full speed and very low boost pressure (under 1 bar) relative to the maximum operating pressure (172 bar). While a pump that operates efficiently with low flow, high inlet pressure, and low boost pressure could not be sourced for the pilot, more efficient pumps are currently used at large scale for circulation in commercial semibatch RO and as booster pumps in conventional RO. The large-scale energy estimates shown in Fig. 7 assume a circulation pump with 7.5% efficiency and a high-pressure pump with 75% efficiency. The pilot required additional components such as a fan, three solenoid valves, and a dosing pump, which we estimate used 0.6 kWh/m3 altogether. In a full-scale system, these additional components would serve a much larger array of membranes with many times the production capacity, so their specific energy consumption is estimated to be lower, on the order of 0.02 kWh/m3.

Fig. 7: Specific energy consumption demonstrated with prototype and predicted at full scale, broken down by the estimated usage of each component.
Fig. 7: Specific energy consumption demonstrated with prototype and predicted at full scale, broken down by the estimated usage of each component.
Full size image

The error bar on the prototype’s total SEC represents meter accuracy.

Figure 7 shows the actual energy use of the pilot system, with an estimated breakdown by component, as well as the predicted energy consumption of a large-scale plant operating on similar feedwater at the same recovery. The large-scale system’s energy model assumed higher-efficiency pumps, lower contributions to SEC from other components, and increased operating pressure relative to the model of ideal mixing. The model predicts that a large-scale system would require approximately 0.8–0.9 kWh/m3 to recover water from low-salinity RO brine, an SEC comparable to a typical potable reuse facility2.

Water chemistry and scaling potential

Table 1 shows the water composition, as measured by the WQIC, of streams entering and leaving the batch RO system. The batch RO system’s concentrate was supersaturated with eight compounds, as revealed by modeling with PHREEQC. Barium sulfate (barite), strontium sulfate, calcium sulfate dihydrate (gypsum), and amorphous silica were considered possible scalants, and their saturation indexes are plotted in Fig. 8, showing very similar scaling potential between the two concentrate samples taken 5 days apart. Fluoride was below the detection limit for both concentrate samples, so it is also possible that calcium fluoride was supersaturated (with a saturation index of up to 1.97). Two crystalline silica species (quartz and chalcedony) were also supersaturated, but were not considered likely scalants relative to amorphous silica because of amorphous silica’s relatively short induction times22. While anhydrous calcium sulfate was supersaturated, its saturation index was lower than gypsum’s, and it does not form spontaneously at room temperature23, so gypsum was considered the more likely calcium sulfate scalant. The low pH of the feed and concentrate prevented supersaturation of alkaline scalants. Good agreement (errors of 0.2% in one sample and 4.9% in the other) between measured and PHREEQC-calculated concentrate conductivity supports the model predictions. Because barite has long induction times24, gypsum was regarded as the primary scalant of concern.

Fig. 8: Saturation indexes (calculated with PHREEQC) of likely scalants in the batch RO system’s concentrate.
Fig. 8: Saturation indexes (calculated with PHREEQC) of likely scalants in the batch RO system’s concentrate.
Full size image

The two samples were taken 5 days apart during the week of continuous operation.

The presence of residual sodium hexametaphosphate (SHMP) in the YDP RO concentrate (the pilot’s feed) likely prolonged induction times and may have helped inhibit scaling. The saturation indexes were calculated in the absence of antiscalant due to the inability to include hexametaphosphate ion in PHREEQC. However, crystalline solids were visible at the bottom of tank storing the pilot system’s concentrate, demonstrating its scale-forming potential despite the residual SHMP.

Despite precipitation in the brine tank, there was no evidence of membrane fouling on the batch RO system’s performance. The system operates at a prescribed flux and was kept at a nearly constant recovery (see Fig. 3), so any fouling significant enough to impact the membrane permeability would have resulted in an increase in operating pressure per the solution–diffusion model25. However, the maximum system pressure did not increase over time (see Fig. 9) beyond an initial settling period which was also reflected in a decrease in permeate conductivity (see Fig. 4). The prototype’s maintenance of permeate quality and recovery without increasing pressure demonstrated the ability of batch RO to concentrate brine to above saturation limits without significant membrane scaling.

Fig. 9
Fig. 9
Full size image

System pressure at the end of each permeation phase.

The week-long operation of the pilot, which produced concentrate at a gypsum SI of 0.63, as shown in Fig. 8, marks a significant departure from our group’s past observations of steady-state RO. A steady-state, plate-and-frame RO system exhibited significant flux decline due to scaling at much lower supersaturations: Flux declined in 13 h with a gypsum SI of 0.16 and in under 3 h with an SI of 0.2626. These observed delays in scaling onset were consistent in order of magnitude with measurements of gypsum’s nucleation induction time at the same supersaturation levels6,27. In contrast, gypsum at an SI of 0.63 has an induction time of approximately 8 min (modeled as in section “Modeling”) in the absence of antiscalants. While the estimated induction time in the concentrate is shorter than the pilot’s permeate production period (about 9 min), the retentate was at a lower concentration throughout most of each cycle and residual SHMP from the primary desalination process was present.

Based on the operating conditions used in the pilot test and the modeling method described in section “Modeling”, gypsum was estimated to progress toward likely nucleation throughout the duration of each cycle as shown in Fig. 10. Using the batch RO system’s measured concentrate composition and the calculated recovery ratio of 82.6%, PHREEQC was used to model the gypsum saturation index (SI) of a range of retentate concentrations. The correlation \({\rm{SI}}=0.5168\ln ({\rm{CR}})-0.2801\), where CR is the ratio of feed volume to retentate volume, was determined from PHREEQC analysis of gypsum SIs based on a feed of YDP concentrate; the correlation’s R2 = 0.9999. (Progress toward barite scaling was similarly modeled and found to be less than 0.01% by the end of a batch due to barite’s very slow nucleation, supporting the assumption that gypsum is the more likely scalant).

Fig. 10
Fig. 10
Full size image

Estimated progress toward gypsum nucleation during a typical permeate production phase.

Figure 10 suggests that the progress toward gypsum nucleation should be incomplete (less than 1/3, where 1 represents likely nucleation), even in the absence of antiscalants, based on the water composition, batch duration and recovery ratio used in pilot testing. The rate of progress toward nucleation rises sharply at the end of the permeate production stage, when concentration rises most rapidly and the nucleation induction time (at high concentration) is the lowest. Given that the nucleation induction time data used to estimate nucleation progress was collected in still solutions of steady concentration and without antiscalants, this estimation method represents only a gauge of the viability of concentrating a feed to a specified recovery ratio with batch RO. This model suggests that it is possible for a batch RO system to avoid scaling, even with a gypsum SI of 0.63 in the concentrate stream, due to the very short time that the RO membrane is in contact with the most concentrated solution. However, further work is needed to refine methods of scaling prediction under conditions of varying supersaturation and determine long-term scaling-free operating conditions in batch RO.

The maximum recovery ratio shown in Fig. 10 (86%), which represents the instantaneous recovery ratio at the end of permeate production, is greater than the overall recovery ratio (82.6%) due to the need to overproduce permeate in anticipation of backwashing during the flushing stage. Producing the extra permeate that is lost to backwashing raises the risk of nucleation; at the same time, backwashing could push unattached crystals or nuclei away from the membrane so that they can be flushed out. Further research is needed to determine the optimal amount of backwashing for long-term scaling resistance.

Qualitative observations of pilot system performance

While the bladder enabled effective batch RO operation, pilot-scale testing revealed difficulties in de-airing of the flexible bladder after installation. It was critical to eliminate air from the bladder to ensure safety because gas is compressible and RO housings are not rated for pressurized gas. After installation, air was removed by manipulating a tube inside the bladder while filling it with water, but it was challenging to determine when all of the air had been removed. Design improvements are needed to enable safe and ergonomic bladder de-airing in larger-scale bladder-batch RO systems.

Testing also highlighted the utility of the independence of flux and recovery in batch RO. In a continuous RO system, changing the operating pressure affects the flux, recovery, and cross-flow velocity at the end of the membrane train. In contrast, in batch RO, flux, cross-flow velocity, and recovery are controlled independently:

  • The speed of the high-pressure pump controls the flux directly because the only fluid outlet during permeate production is the permeate itself

  • The speed of the circulation pump controls the cross-flow velocity, which varied only about 25% throughout the 2-membrane train used in this system

  • The batch RO system’s recovery is determined by the ratio of the bladder volume to the feed volume brought in each cycle, so recovery can be set on a cycle-by-cycle basis

Independent tuning of flux, cross-flow velocity, and recovery ratio was useful in a pilot setting. In a full-scale system, it could allow for operation that adapts rapidly to changes in feed composition, demand for water, or the price of electricity.

Discussion

While the batch RO pilot desalinated RO concentrate without scaling, the conclusions we can draw are limited by the short testing time. Although the pilot completed >800 cycles, a week-long test is not sufficient to prove the absence of long-term scaling processes that might occur over many weeks or months. For example, if there is a low but nonzero likelihood of crystal nucleation and deposition on the membranes in each cycle and the backwash step is not adequate to remove such newly-formed crystals, scaling could accumulate during long-term operation. Furthermore, long-term effects of rapid pressure cycling on critical components such as O-rings could appear later. Longer-term batch RO pilot studies are being conducted to determine whether there are delayed effects of scaling or pressure cycling.

While system performance stayed steady for the week-long investigation, the absence of performance decline cannot prove the absence of scale formation. Furthermore, scaling likelihood was impacted by residual SHMP in the pilot’s feedwater, so it is unknown whether scaling would have occurred in a conventional RO process with the same recovery ratio. Another study is underway to directly compare batch RO and conventional RO.

The prototype developed for this study was small in scale, with only a single bladder pressure vessel. While it produced high-quality permeate, it could not attain the low specific energy consumption estimated at full scale due to the low efficiency of small-scale components and the inability to spread the energy consumption of the control system (including valves) across many membranes. An additional pilot study is in progress to explore batch RO performance at larger scales.

While it is important to consider the limitations discussed above, the batch RO pilot successfully recovered water from a real RO concentrate. In a week of continuous testing, comprising 885 batch cycles, the batch RO system:

  • Recovered 82.6% of the concentrate it received.

  • Produced permeate with an average salinity of 150 ppm TDS.

  • Used an average of 633 W of power for a specific energy consumption of 3.3 kWh/m3, which could potentially be reduced to under 1 kWh/m3 in a full-scale system.

  • Produced concentrate likely supersaturated with gypsum, barium sulfate, strontium sulfate, and amorphous silica without any performance deterioration.

The findings of this pilot-scale study demonstrate the potential of batch RO for high-recovery, low-energy desalination and water reuse. Further investigation is needed to explore batch RO’s long-term fouling resistance and energy consumption at larger scales.

Methods

A small, pilot-scale prototype of batch RO with a flexible bladder was constructed and used to desalinate RO concentrate from the Yuma Desalting Plant (YDP) in Yuma, Arizona, USA. After initial testing in a laboratory with sodium chloride solutions (reported in ref. 7) and a period of shakedown testing at the YDP, the pilot system was operated continuously for 1 week while flow rates, permeate composition, and energy consumption were monitored.

The YDP uses RO to desalinate saline agricultural drainage water from the nearby Main Outlet Drain Extension. The reject stream from the YDP’s primary desalination process was used as the feed to the batch RO pilot. This stream fed to the pilot had, on average, 2521 mg/L sulfate and 6366 mg/L total dissolved solids during the weeks surrounding the pilot test. (Refer to Table 1 for detailed water composition.) The YDP reject stream fed to the batch RO system contained approximately 5.6 mg/L of the antiscalant sodium hexametaphosphate (SHMP), which was estimated based on YDP feed concentration of 1.5 mg/mL SHMP, as communicated by email from Amy Klein (October 19, 2021) and estimated YDP recovery. Sulfuric acid (0.1 M) was injected occasionally to reduce the pH and minimize the risk of calcium carbonate scaling. Using the YDP’s sulfate-rich concentrate as the pilot system’s feed allowed for pilot testing with a real, scaling prone-water source.

The pilot batch RO system, pictured in Fig. 11, was designed to produce up to 5 m3/d and have a maximum system pressure of 34.5 bar. Through the use of a flexible bladder inside a standard RO housing, the design enabled realization of a true batch process—in which no feedwater entered the system during permeate production—using almost exclusively off-the-shelf equipment. The physical system and the software used to control it are discussed in this section.

Fig. 11
Fig. 11
Full size image

Photograph of the pilot system in operation at the Yuma Desalting Plant labeled with key components.

The flexible bladder, which separates a single pressure vessel into two compartments, allows for constant pressurization of the circulated retentate as its volume shrinks due to permeate production13,28. As sketched in Fig. 2, the bladder was installed inside an RO housing, turning a standard RO component into a variable-volume, high-pressure tank. As the bladder is filled with pressurized tap water, it compresses the surrounding volume of retentate; when the cycle is finished, the tap water is returned to a tank for use in the next cycle. The bladder was selected from among designs previously proposed because it can be constructed of inexpensive materials and has no sliding parts that could wear during long-term operation. The bladder also isolates the high-pressure pump from the feed, which is typically a corrosive liquid such as RO concentrate. The high-pressure pump only contacts tap water (or another liquid of choice), and thus can be constructed of inexpensive materials and expected to last longer.

The risk of fouling can be reduced by frequent osmotic backwashing of the RO membranes29,30, which occurs between cycles in this batch RO system. After each 10-min permeate production phase, the system is depressurized in order to flush the remaining brine, empty the bladder, and refill the system with fresh feed. At the onset of this reset process, a sudden flow of product water back into the system results from osmotic backwash across the membranes. Although this slightly reduces the overall water recovery, osmotic backwashing exposes the membrane to subsaturated water and fluid flow away from the membrane immediately after the period with the highest risk of crystal nucleation31.

Pilot system design

Figure 12 shows the system design as a process flow diagram. Saltwater (referred to here as the retentate) circulates at high pressure around the retentate loop, which includes the space between the bladder and bladder housing (an 8”, or 20 cm, RO housing). During permeate production, all valves (whether normally open or closed) are un-powered, and a pressurizing liquid is pumped into the flexible bladder at high pressure. In the pilot test, the pressurizing liquid was a mixture of tap water and YDP RO concentrate, but the composition of the makeup liquid does not matter as long as it is compatible with the pump materials because it does not contact the water being desalinated. As liquid is pumped into the bladder, pressure equalizes across the flexible bladder, thus pressurizing the retentate. Water is nearly incompressible, so as water is pumped into the bladder, the volume displaced by the growing bladder exits the system through the RO membrane (Toray TM710D) as permeate. As permeate leaves the retentate loop, the retentate becomes increasingly saline, and its osmotic pressure increases. The retentate loop pressure rises to the value needed to enable permeation to occur at the flux mandated by the pressurizing pump’s flow rate per unit membrane area. No fluid is introduced to the retentate loop during permeate production.

Fig. 12: Process flow diagram of pilot system with the main flow loops highlighted.
Fig. 12: Process flow diagram of pilot system with the main flow loops highlighted.
Full size image

“N.C.” denotes a normally closed solenoid valve while “N.O.” denotes normally open.

After each permeate production stage is complete, brine is flushed from the retentate loop by an influx of feed (i.e., YDP RO concentrate), which was delivered at a gauge pressure of about 2 bar. During the flush phase, all valves (whether normally open or closed) are powered, and feed flows into the retentate loop, pushing the brine out of the system. Then, during the recharge phase, only the valves in the pressurizing loop are powered; no brine leaves the system while the feed fills the space around the bladder, pushing the pressurizing water back into the open tank for use in the next cycle. Both pressurizing and circulation pumps (both Wanner Hydra-Cell reciprocating models) were controlled by variable frequency drives. The peristaltic acid-dosing pump, used to reduce pH as needed by injecting 0.1 M sulfuric acid, was powered only during the flush and recharge phases. During flush and recharge, the system pressure was low (0–2 bar gauge), so permeate would flow backward through the membrane, providing an osmotic backwash. The accumulator installed in the permeate line (in the prototype, this was simply an empty cartridge filter housing) ensured there was adequate permeate available so that the membranes did not dry out due to backwashing and permeate loss could be continuously measured. Although permeate flows back through the membrane during the flush and recharge stages, neither permeate flowmeter could detect changes in flow direction. A system of four check valves (omitted from Fig. 12 for clarity but described in the Supplementary Information) was used to direct flow through the first permeate flowmeter in its preferred direction regardless of whether permeate was being produced or back-flushing; the software assigned a sign to the flow rate value based on the stage of operation.

Relative to past designs for bladder-batch RO8 this design uses fewer actuated valves (three 2-way solenoid valves instead of five). Instead, it utilizes inexpensive check valves where shown in Fig. 12.

The bladder used in this prototype is shown in Fig. 13. It was custom-designed to fill the inside of an 8” (20 cm) RO housing (Pentair Codeline), as shown in Fig. 13c. It was sewn from thermoplastic polyurethane–coated 200-Denier Oxford-weave nylon (Seattle Fabrics) as shown in Fig. 13b. The seams were sealed with a household iron and, at corners, a polyurethane-compatible adhesive (Seam Grip). The bladder was clamped onto a plastic cylinder studded with holes (modeled after an RO membrane’s permeate tube), which was then installed in the RO housing like a spiral-wound membrane. RO feed spacers were wrapped around the cylinder to prevent the bladder fabric from blocking the holes, and rubber tubing was placed between the tube and the bladder to create a tight seal when clamped. A plastic rib was also built into the spacer to stiffen it and facilitate installation. As shown in Fig. 13a, a polyurethane tube, capped with a check valve, passed through a waterproof cord grip to facilitate air removal from the bladder (a safety-critical step performed during bladder installation).

Fig. 13: Photographs of the bladder design.
Fig. 13: Photographs of the bladder design.
Full size image

a The assembled bladder, showing de-airing valve; (b) detail of bladder construction; (c) bladder (a nylon prototype) filled with water in an open RO housing to demonstrate its operation.

A few additions to a standard RO housing were needed to facilitate bladder use. A blind adapter, visible in Fig. 13c, was used to internally block one permeate port of the bladder housing. Due to the low pressure rating of standard permeate ports, the port fed by the high pressure pump was replaced by a 316 stainless steel permeate port with a pressure rating of 600 psi (41.4 bar) in a metal head assembly. Short plastic tubes with several holes drilled in the sides were pressed into the inside of the pressure vessel’s side ports to prevent the bladder from sealing against the side ports and blocking flow. A long plastic tube was used to connect these two short tubes to maintain a flow channel around the bladder even when it was very close to full. These last two modifications were made after observing a failure mode in which the bladder was pulled into the outlet side port.

The pilot was instrumented with devices for measurement and control as shown in Fig. 12, including feed and permeate flow rate sensors, a pressure transmitter, and a differential pressure transmitter. A data acquisition unit (National Instruments) was used to transmit signals between instrumentation, the laptop used for control, and the pumps and solenoid valves.

Some flow, energy, and water composition data pertaining to the pilot system was independently collected by the pilot site. This data was provided by email from Amy Klein on behalf of Carrot on August 19 and September 12, 2022 as part of the More Water Less Concentrate Prize Competition. Feed, permeate, and concentrate flow rate sensors were installed by the U.S. Bureau of Reclamation’s Water Quality Improvement Center (WQIC) staff to collect data on the flows into and out of the pilot system. Daily energy use was also recorded by the WQIC. These data were used in analyzing the pilot’s recovery ratio, permeate production rate, and energy consumption. The WQIC also provided laboratory measurements of constituent concentrations, pH, and total dissolved solids of samples of the pilot system’s feed, permeate and concentrate.

System control

The pilot was controlled by a MATLAB script running on a Windows laptop. The script performed three principal functions: (1) graphical user interface; (2) communication with data acquisition unit; and (3) control logic.

The graphical user interface allowed users to both control the batch RO system and to continuously monitor system status and sensor readings. A manual operation mode allowed the user unilateral control to switch between defined phases for system shakedown and debugging. Switching to automatic mode engaged the control logic (outlined below) to automatically and repeatedly run the batch RO cycle.

The MATLAB script interfaced with the data acquisition unit, reading sensor data through analog input channels and actuating valves through the digital output channels. The high-pressure pump ran at a fixed speed during permeate production, causing approximately constant flow rate into the bladder and thus also flux through the membrane; pressure was not actively controlled, so control consisted of switching between phases of operation at the right times.

The control logic continuously monitored sensor data to determine whether it should exit the current phase and move on to the next as follows:

  • Permeate production ended when the bladder was full, which was determined when the pressure difference between side ports of the bladder housing exceeded a certain level (e.g., 0.5 bar). Alternatively, permeate production would end early if the system reached maximum absolute pressure.

  • Flush ended when the target volume of brine (e.g., 5 L) had been flushed out by the entering feed.

  • Recharge ended when the bladder was nearly empty, which was determined by when the feed flow rate fell below a certain level (e.g., 20 L/min).

In addition to the above primary exit conditions, the control logic included back-up exit conditions based on time or volume in case of sensor failure or other unexpected behavior.

Modeling

To estimate scaling likelihood, the retentate concentration was modeled using the batch RO model of ref. 15. This model discretizes the feed channel in space and time and accounts for concentration polarization. The model predicts the retentate concentration both in the bulk solution and at the membrane as a function of time throughout the permeate production stage. The predicted retentate concentration as a function of time then permits estimation of scaling likelihood and energy consumption.

A new model was developed to estimate nucleation likelihood in the batch RO system based on the kinetics of scaling under time-variant conditions. The model builds on the work of Warsinger et al.6, which used a recovery ratio–based brine concentration to determine the minimum induction time of likely scalants and thus determine safe cycle durations in batch RO. The new model, described below, takes a step toward incorporating time-varying conditions into a kinetic model of scaling. While it lacks extensive validation, the model successfully predicted a scaling-free recovery setpoint during pilot testing of this batch RO system.

The saturation indices of possible scalants were predicted using PHREEQC32 based on water compositions measured by the WQIC. For both typical YDP concentrate composition and the measured composition of the batch RO system’s concentrate, and as described in section “Water chemistry and scaling potential”), the most likely scalant in the batch RO pilot system was gypsum (calcium sulfate dihydrate).

The maximum retentate concentration at each timestep (modeled using ref. 15), which occurred at the membrane surface near the end of the second membrane, was used to calculate the maximum gypsum saturation index within the batch RO system as a function of time. Saturation index is the base-10 log of the ratio of ion activity product to solubility product. Using a correlation between gypsum saturation index and induction time from ref. 6 based on experimental induction time data in ref. 27, the rate of progress toward nucleation was calculated as the inverse of induction time. Possible effects of turbidity (shown to hasten scaling33) and antiscalants were neglected in this preliminary model. Nucleation progress rate was numerically integrated from the beginning of the permeate production phase to estimate the nucleation progress Φ at a given time t in each batch cycle, as described in Eq. (3). Backwashing during the flushing stage, as modeled in ref. 31, was assumed to flush away any highly supersaturated solutions or nuclei, thus resetting the nucleation progress to zero at the beginning of each cycle.

$$\Phi (t)=\mathop{\int}\nolimits_{0}^{t}\frac{1}{{t}_{ind}({\rm{SI}}({t}^{{\prime} }))}\,{\rm{d}}{t}^{{\prime} },$$
(3)

where tind is the induction time associated with the current saturation index, SI, of a potential scalant.

In addition, energy consumption was modeled for both the pilot and a hypothetical large-scale batch RO system. Although the Swaminathan et al. model15 was developed for a batch RO system with an ambient-pressure tank and a pressure exchanger, this study’s pilot system uses a bladder rather than a pressure exchanger for retentate pressurization. The Swaminathan et al. model of energy consumption was used for this bladder-based system by eliminating pressure exchanger-related losses, i.e., using a value of 1 for pressure exchanger efficiency.