1. Introduction
Water scarcity is a major problem nowadays in both coastal and inland areas [
1]. Desalination of sea and brackish water has become the most widespread solution to alleviate this issue [
2]. Membrane technologies are the most used at industrial scale. Reverse osmosis (RO) is the most well-assessed technology for decades. Results obtained at both laboratory and industrial scale have allowed the development of new membranes and modules, and also numerous performance and costs models. All this experience throughout the time leads to RO being the most economically competitive sea- and brackish water desalination operation so far [
3,
4]. RO plants produce 70% of the total permeate worldwide, followed by thermal desalination operations such as multi-stage flash (MSF) and multi-effect distillation (MED) [
5], with 16% and 7%, respectively. However, with the increasing freshwater demand worldwide, the disposal of concentrated brines to the sea or to wells could be an environmental problem with major relevance in the near future [
6]. This fact brings to light the need of alternative desalination technologies able to deal with the concentrated brines that RO cannot handle because of the increased osmotic pressure [
7,
8]. Apart from producing more freshwater, a proper brine management will also give the opportunity of recovering valuable minerals from that residue and thus establish a circular economy scheme with zero liquid discharge (ZLD) as the final goal [
9,
10,
11]. Among these novel brine concentration technologies, membrane distillation (MD) is proposed to supplement RO [
9,
12].
MD is a thermal desalination technique implemented at pilot scale for two decades and driven by the vapor pressure gradient between both sides of a hydrophobic microporous membrane, not by osmotic pressure [
13]. This fact makes possible the treatment of brines that cannot be handled by RO due to their high salinity. However, the main issue reported in high-salinity MD is pore wetting induced by salt ions, which occurs when the hydrophobicity of the membrane pores (related to the diameter of the pores and the surface tension of the feed) is reduced, facilitating thus the pass of salts through the membrane and hence worsening the permeate quality [
14]. Moreover, in the worst case, crystal growth can occur and cause severe irreversible membrane damage [
11,
15]. Membranes made of different materials and with different structures have been tested at a laboratory scale [
16,
17,
18]. However, only a few studies have assessed the performance of MD at high salinity up to date. Several studies about the treatment of brines with concentrations up to 240 g L
and with membrane areas below 0.02 m
have been published, considering the MD operational modes direct contact (DCMD) [
19,
20], air gap (AGMD) [
21,
22], and vacuum (VMD) [
23]. A minor upscaling up to almost 0.2 m
membrane area was later made. Valuable information about the treatment of real RO brine was provided in these studies carried out in DCMD [
24] and VMD [
25] operational modes. However, these results at a lab scale can hardly be extrapolated to bigger MD modules, therefore experimental evaluations in commercial-scale devices are mandatory to cover the current lack of knowledge [
18].
One of the first assessments of pilot-scale MD modules in which permeate quality was considered in detail was published by Minier-Matar et al. [
26]. Permeate electrical conductivity of <10
S cm
was measured in the treatment of brine with concentration 70 g L
using two commercial plate-and-frame vacuum multi-effect MD (VMEMD) modules with membrane areas of 6.4 and 4.6 m
. This value is in the range of those obtained by Andrés-Mañas et al. [
27] in a similar 6.4-m
VMEMD module but operated with real Mediterranean seawater, which means a salt rejection factor (SRF) of above 99.98%. Regarding pilot-scale spiral-wound modules, values of permeate electrical conductivity up to 370
S cm
(two orders of magnitude higher than those of lab-scale studies) were reported.
These results can be explained by the fact that the membrane pore diameter follows a Gaussian distribution, and hence as the membrane size increases, so does the number of pores with excessive size to maintain the hydrophobicity, increasing the liquid flow through the membrane. Winter et al. [
28] evaluated a permeate gap (PGMD) single-envelope module with a membrane area of 10 m
under feed salinities between 0 and 100 g L
, but their work was not focused on permeate quality, although they presented values of permeate electrical conductivity. Soon after, Ruiz-Aguirre et al. [
29] carried out a thorough analysis of the permeate electrical conductivity along the treatment time of a feed with 35 g L
marine salts in a similar PGMD module. The authors demonstrated that in the very beginning of the operation, the permeate itself acts rinsing the permeate channel, removing the liquid feed that has leaked through the membrane pores by microfiltration during the stand-by period, and therefore the permeate quality improved with time and stabilized below 20
S cm
. An alternative method to rinse the gap and maintain good permeate quality in a longer term was proposed by Schwantes et al. in a bench-scale single-envelope air-gap (AGMD) module [
15]. Blowing air into the gap increased the absolute pressure on it, which is detrimental for vapor diffusion through the pores. The final result was a significant reduction of the permeate electrical conductivity when treating hypersaline feeds with up to 240 g kg
regarding the same tests performed without air sparging.
The subsequent development of multi-envelope AGMD modules increased the performance of the MD since higher feed flow rates with larger membrane area could be treated without an excessive hydraulic pressure drop inside the module [
30]. This improvement in permeate flux and specific thermal consumption was characterized by Ruiz-Aguirre et al. [
31]. Two multi-envelope modules named AS7 and AS24 were used, comprising 12 internal channels, but of different lengths (1.5 m in the former and 5 m in the latter). A multi-objective optimization of the trade-off between permeate flux and heat recovery was proposed for each one, but without taking into account the permeate quality.
Up to date, air suction from the gap and the membrane pores of multi-envelope AGMD modules has demonstrated to be the most successful solution to increase the MD performance. The vacuum-assisted air-gap MD (V-AGMD) operation is based on reducing the absolute pressure in the gap enough to improve vapor diffusion through the membrane pores, but without affecting the condensation inside the module. Values of permeate productivity reported in a module AS7 were up to 8.7 L h
m
, similar to those obtained with the vacuum multi-effect technology (VMEMD). Besides, permeate electrical conductivity figures were in the same range (below 50
S cm
) [
32], although the effect of vacuum on permeate quality must be thoroughly evaluated, especially in brine concentration processes. The only detailed study so far on this topic in pilot-scale modules was published by Ruiz-Aguirre et al. [
33]. The authors showed that the SRF values of the modules AS7 and AS24 were worsened up to 2% by increasing the feed salinity up to 140 g L
, and up to 1.5% when the absolute pressure in the gap was about 200 mbar, compared to experiments carried out in AGMD mode. On the contrary, the permeate quality in the single-envelope PGMD module remained almost unchanged and close to 100%. This suggests that the combined effect of high salinity and vacuum promotes membrane wetting. To quantify it, a performance parameter named membrane leak ratio (MLR) was introduced by the authors in the same study, which is defined as the ratio of feed that passes through the membrane pores in operation. In the worst case of SRF reported (97.2%), values of MLR < 0.12% were calculated, which brings to light the extreme dependence of permeate quality on the hydrophobicity of the membrane pores.
Subsequently, the performance of three V-AGMD modules AS7, AS24, and AS26 was compared, having the latter twice as many channels as the AS24 but with around half the length, which also reduces the circulation velocity by half, and thus the hydraulic pressure inside the channels. Experiments showed that the module AS26 outperformed the AS7 and the AS24, and provided the lowest specific thermal consumption reported to date in the MD literature: 40 kWh
m
, equivalent to GOR = 16.4. Therefore, the multi-envelope module AS26 is the strongest candidate to be part of a potential upscaled MD facility competitive with those of other technologies [
34]. The authors provided SRF results higher than 98.2% and maximum membrane leak ratio (MLR) of 0.19% for several operating conditions, but no conclusive modeling of the effect of each variable on these quality parameters was performed on the module AS26.
Since permeate quality is subject to unknown defects on the membrane, the use of artificial neural networks is justified for modeling under different operating conditions [
35,
36]. Artificial Neural Networks (ANN) have emerged as a promising modeling tool in the realm of MD systems, especially for flux and thermal efficiency prediction [
37] and fouling prediction [
38,
39]. One of the primary advantages of this methodology is its capability to effectively capture and fit almost all nonlinear processes. Furthermore, the inherent structure of the model enables retraining with additional experimental data, offering the potential for further enhancing prediction accuracy.
Multi-envelope modules operated in V-AGMD mode have been fully characterized due to their improved results. Concretely, the module AS26 has been modeled in depth in terms of heat recovery and permeate productivity, and operating conditions that optimize the performance have been established [
40]. However, there are no studies focused specifically on assessing and modeling the permeate quality of the operation. This work presents for the first time the application of ANN to model and simulate the permeate quality of a pilot-scale MD system as a function of the operating conditions, providing a machine-learning framework for the application at hand. To do that, a comprehensive experimental campaign has been carried out at the solar MD facilities of CIEMAT-Plataforma Solar de Almería, using a pilot-scale multi-envelope module AS26 operated in V-AGMD mode. Outputs of the model are the key quality indicators salt rejection factor (SRF) and membrane leak ratio (MLR), whereas inputs are the operating conditions: evaporator channels inlet temperature (TEI), cooling channels inlet temperature (TCI), feed flow rate (FFR), and feed salinity (S). Finally, operating limits have been established, and techno-economical aspects of the operation in a wide range of conditions are discussed, focusing mainly on how both the permeate quality and the thermal efficiency are affected.
4. Conclusions
V-AGMD operation on a commercial multi-envelope air gap module AS26 has been described in previous studies as the most energy efficient up to date. This makes MD technology a great candidate for the valorization of concentrates from other desalination operations that are less tolerant of high salinity, such as RO. However, no detailed studies about permeate quality in V-AGMD with this module and considering the feed salinity as a variable have been carried out so far. The present study is intended to fill the lack of information in the MD literature on this topic.
The experimental campaign comprised tests with different setpoints of inlet temperatures, feed salinity, and feed flow rate. With these results, two performance metrics that allow quantifying the quality of the permeate were calculated: SRF and MLR. The variables TEI, FFR, and S had a very noticeable influence on these two metrics, whereas TCI had a negligible effect.
Owing to the non-linearity of the experimental results, machine learning techniques were applied to characterize the two responses in relation to the aforementioned inputs, rather than common regression methods. Thus, the SRF and the MLR were modeled with an artificial neural network comprising two hidden layers with 4 and 2 perceptrons, respectively. Model validity was demonstrated by R-values close to one in the train (0.95), validation (0.89), and test (0.97) datasets.
Analysis of the experimental results showed two competing effects on permeate quality: the dilution of the feed leaked through the membrane pores and the hydraulic pressure. Under operating conditions that favor the driving force, i.e., TEI higher than 70 C, FFR higher than 750 L h and S up to 175.3 g L, their influence on permeate quality was little. Thus, the SRF values were within 99.99% with S = 35 g L and 98.54% with S = 175.3 g L. These are equivalent to MLRs among 0.007% and 0.084%, respectively. In these cases, the vapor flux through the membrane pores is much larger than that of the feed leaked through the pinholes, so the dilution effect of the leak through the potential water doors allows maintaining high quality in the permeate. However, from an energy point of view, it must be taken into account that, in this salinity range, operating costs are optimized by reducing the feed flow rate.
When operating conditions hinder a high vapor diffusion through the membrane pores, the effect of hydraulic pressure surpasses that of leaked feed dilution. For feed salinity up to 70.1 g L, working with low TEI and FFR had a negative effect on both the SRF and the MLR, since the dilution effect diminished. However, at higher feed salinity, increasing productivity meant improving the SRF by up to 96%, but also worsening the MLR by up to 0.2%, since the reduced volume of permeate did not balance out the effect of the hydraulic pressure in the channels, together with the reduced hydrophobicity of the membrane due to the high salt load in the feed. Despite the aforementioned permeate quality values, no evidence of irreversible wetting or scaling was observed during the experimental campaign because the percentage of leaked feed in operation was very low.
This study has demonstrated the great tolerance to salinity of the pilot-scale V-AGMD module AS26 and the great hydrophobicity of the membranes currently used, which maintain their integrity even under conditions of high hydraulic pressure within the module channels, without signs of wetting or scaling. The main drawback identified was that the salinity requirement demanded by the WHO for drinking water (0.35 g L) was only met in the treatment of feed with marine concentration, although the separation of salts was very efficient.
In conclusion, considering both the thermal and the permeate quality performance, the V-AGMD module AS26 can be competitive for brine concentration. In any case, if the production of drinking water is seeked when using the modules for brine concentration, improved salt rejection by using superhydrophobic membranes must be considered in future works to maintain the permeate quality in a wider range of feed concentrations.