**1. Introduction**

Particulate systems like microparticles have attracted interest in several biomedical, food and environmental applications [1–5]. Namely, the encapsulation of bioactive agents in these carriers improves their efficacy and safety, since better control of the dosage and release are provided [6,7]. Microparticles also enhance physicochemical stability, protecting the cargo from environmental and physiological factors [8]. The size of microparticles, between 0.1–100 μm [9], can hamper their absorption through biological membranes, increasing their permanence in the application site, thus providing local and sustained drug release and mitigating their toxic effects [10].

Lipids are advantageous matrices for particulate drug delivery systems since they are physiological compounds and therefore well tolerated by living systems [11,12]. For instance, a variety of lipids such as sorbitan esters, phosphatidylcholine, and unsaturated polyglycolized glycerides are widely used as surfactants in lipid-based formulations [13]. Among lipid systems, solid lipid microparticles (SLMPs) are easy to produce on a large scale and sterilize, exhibiting better stability properties than others, such as liposomes [14]. Several SLMP-based formulations have been developed as drug delivery systems for oral, parenteral, pulmonary and topical applications [14,15].

Solvent-free strategies are especially attractive for the manufacturing of SLMPs from the processing, environmental and economical points of view. Namely, supercritical CO2 (scCO2) technology has been highlighted as a processing tool for environmentally friendly, safe and cost-efficient techniques at mild conditions—pressure (P) > 73.8 bar and temperature (T) > 31.1 ◦C) [16]. Processes based on supercritical fluid technology (foaming, sterilization) usually avoid or at least mitigate the use of organic solvents thus reducing their carbon footprint. The PGSS® (Particles from Gas-Saturated Solutions) technique is based on the use of compressed CO2 or scCO2 for the production of microparticles in an atomization-wise process [17–19]. PGSS® process comprises two main steps: (i) CO2 sorption in the polymer, and (ii) polymer expansion and particle formation. In the first step, high amounts (5–50 wt.%) of CO2 dissolve in a molten substance at a moderate pressure in an extent depending on the soaking time and CO2 affinity to the polymer [20]. Then a rapid expansion to atmospheric pressure of the melt through a nozzle causes an intense cooling effect and CO2 supersaturation within the melt, resulting in the precipitation of solid particles [21]. scCO2 used in the PGSS® technique differs from other compressed fluids (e.g., compressed air) used in conventional atomization processes (spray drying) in their chemical interaction with the processed polymers at a molecular level, as scCO2 can decrease the melting temperature of the polymer thus contributing to costs optimization and energy consumption savings [22]. PGSS® is an adequate technique for the processing of polymeric particles incorporating thermolabile compounds, although its use is limited to polymer matrices with relatively low melting temperatures and with an affinity of CO2 to the polymer [23]. Compared to other processes for particle production involving the use of scCO2, such as the gas antisolvent (GAS), supercritical antisolvent (SAS) and supercritical fluid extraction of an emulsion (SFEE) techniques, the PGSS® technique does not use any organic solvents [16,24]. Moreover, the substance to be micronized does not require to be soluble in CO2 unlike in the rapid expansion of supercritical fluids (RESS) process [25,26]. Overall, PGSS® emerges as an appealing and advantageous technique for the processing of SLMPs at reduced melting temperatures and in the absence of organic solvents.

The morphology and size of the SLMPs produced by the PGSS® process are mainly influenced by the formulation (chemical composition and rheology of the compounds to be precipitated), the technical details of the equipment used (volume of the saturator, precipitator and collector, diameter of the nozzle and length of the tubing) and the operating conditions (pressure, temperature, soaking time) [27,28]. The PGSS® processing variables are numerous, making it difficult to elucidate their influence on the characteristics of the microparticles using conventional statistical methods [29–31]. Despite PGSS® being a simple and versatile method, the lack of knowledge of the effects of the variables on the results of PGSS® technology may entail an obstacle towards the robust SLMPs production and the scaling-up of the process [32]. Approaches based on DoE (design of experiments) and multiple regression have been proposed to manage the number of experiments, to select the critical variables and to optimize the operation conditions, but mainly regarding their influence on the dissolution profile of the drug incorporated in the particles [33]. Some mathematical models were also proposed to simulate the physicochemical processes taking place during the PGSS® processing, such as the behavior of a CO2-supersaturated solution drop in low-pressure environments [34,35]. In this context, artificial intelligence technologies emerge as tools with grea<sup>t</sup> potential for simplifying the study of processes in which many variables are involved, even when a small number of experiments are available. Some of them, such as the neurofuzzylogic systems, allow multiple variables to be modeled and the models expressed through language, which generates in-depth knowledge about the process. Neurofuzzylogic software is a hybrid system that combines artificial neural networks (ANN) and fuzzy logic (FL). ANN are computer programs that simulate how the human brain processes information. They detect patterns and relationship in data, and learn from experience, leading to "black-box" mathematical models [36]. When combined with FL, the models are expressed as simple linguistic IF ... THEN rules together with a membership degree, losing their black-box character and being easily understandable.

Artificial intelligence tools have been previously used in the development and optimization of microparticles [37] and polymeric and lipid nanoparticles [38,39]. To the best of our knowledge, these tools are applied in this work for the first time to model the production of SLMPs by the PGSS ® technology. SLMPs consist of a matrix of commercial glyceryl monostearate (GMS), a lipid widely used as an emulsifier in pharmaceutical preparations due to its good biocompatibility and safety [40,41]. First, the melting point depression of commercial GMS in contact with scCO2 was studied to establish the limits of the adequate knowledge space for the processing of PGSS ®. Subsequently, an unconstrained D-optimal design for three variables (nozzle diameter, pressure and temperature) at 2, 3 and 3 levels, respectively, was used to prepare SLMPs using the PGSS ® technique. The microparticles were characterized in terms of size and shape. The generated database was modeled through a neurofuzzylogic system and the design space was established with respect to the melt GMS processability (fine particle production yield) and the characteristics of the particles.

### **2. Results and Discussion**

### *2.1. Melting Point Depression of GMS in the Presence of CO2*

Melting pressure-temperature curve of the commercial GMS under compressed CO2 was measured to determine the feasible operating range of conditions for the PGSS ® technique (Figure 1). This step is crucial since it is necessary to establish a set of pressure-temperature conditions (grey region in Figure 1) where the lipid mixture is molten. The melting point of GMS in the presence of CO2 has been previously studied [42], but these determinations are essential because it is well known that GMS can have inter-batch and inter-manufacturer variability as it is commercially provided as a mixture of components (mono- and diglycerides).

**Figure 1.** Glyceryl monostearate (GMS) melting points obtained at di fferent pressures of CO2 using a variable-volume high-pressure view cell. Grey area represents the pressure-temperature region at which GMS will be molten. The area delimited by the dashed line represents the operating region established for solid lipid microparticles (SLMPs) production by PGSS ® technique.

The melting point of the commercial GMS without CO2 was 61 ◦C at ambient pressure. CO2 can act as a plasticizer agent, being able to melt other substances, like lipids or polymers, below their normal melting points. Melting point depletion e ffect of GMS in contact with CO2 is highly dependent on the working pressure and decreased proportionally up to 52 ◦C as can be seen in Figure 1. This e ffect was related to the increase in the amount of CO2 dissolved in the lipid when the pressure increases [43]. A plateau in temperature was reached at 52 ◦C and pressures above 120 bar were not able to cause an additional melting point depletion. This second e ffect was related to the competing mechanism of increased CO2 solubility in the lipid and the hydrostatic pressure promoting the melting point depletion and increase, respectively, that are counteracting at pressures above 120 bar for GMS [42]. The reduced melting temperature in the presence of compressed CO2 is advantageous for the energy optimization of the PGSS ® particle processing when transferring formulations from lab to pilot scale [44,45].

### *2.2. Particle Size Distribution (PSD), Morphological and Physichochemical Characterization of GMS Particles*

Based on the melting point values obtained in Section 2.1, the range of values of pressure and temperature selected for the experimental study of the PGSS ® processing of GMS particles were set at 120–200 bar and 57–67 ◦C, respectively. In this work, an increment of ca. 5 ◦C with respect to the melting temperature of GMS at a certain pressure in the presence of compressed CO2 was established as a rule-of-thumb (dashed and grey rectangle in Figure 1) to ensure the complete melting and to avoid clogging of the nozzle during the PGSS ® expansion-spraying step. The selection of the nozzle diameter was based on the technical possibilities of the PGSS ® equipment, being 4 and 1 mm the maximum nozzle diameter and the minimum nozzle diameter that did not cause clogging events upon depressurization using the established P-T range in the experimental design, respectively.

PSDs of the SLMPs showed mean diameters between 100 and 190 μm and standard deviations between 30 and 65 μm (Table 1). In general, the PSDs fitted well to a normal distribution (Figure 2) with good correlation levels (R<sup>2</sup> > 0.95) in all cases. The yield of particle production was determined from the weight percentage of fine particles with respect to the initial GMS (Table 1). The loss of material during the PGSS ® processing was due to GMS remaining in the tubing and the saturator of the equipment, molten material that was not solidified into particles and formed a crust in the walls of the precipitator. Some mass losses were attributed to small particles that remained suspended in the outlet gaseous stream and were vented out during the depressurization step along with the CO2.


**Table 1.** Yield of particle production, mean diameter and standard deviation of SLMPs of GMS processed using PGSS ® technique. Particles were denoted as GMS-x-y-z, where x is the nozzle diameter (mm), y the processing temperature (degrees Celsius) and z the processing pressure (bar).

**Figure 2.** Frequency histogram of GMS-1-67-200 particles (mean particle diameter = 125.4 ± 43.1 μm). The normal distribution of this histogram is representative of all the GMS formulations tested.

The processing using PGSS® technique led to particles with reduced circularity (60.7 ± 18.2%) with respect to the original GMS (round particles, Figure 3A). PGSS®-processe<sup>d</sup> lipid microparticles had a decreased bulk density (0.14 g/cm3) with respect to the raw material (0.53 g/cm3). However, skeletal density was similar (0.995 ± 0.017 g/cm3) to the unprocessed GMS (0.980 ± 0.003 g/cm3), suggesting that the chemical structure of the GMS was not unaltered during the process, as also confirmed by X-ray diffraction (XRD) and Attenuated Total Reflectance/Fourier Transform infrared spectroscopy (ATR/FT-IR) (Figure A1).

**Figure 3.** Effect of temperature in the PGSS® processing of GMS particles: (**A**) unprocessed GMS particles and (**B**) GMS-1-57-200, (**C**) GMS-1-62-200 and (**D**) GMS-1-67-200 particles.

*2.3. Morphological Characterization and Modeling of GMS Particle Production Using Neurofuzzy Tool*

The processing of GMS using the PGSS® technique resulted in porous particles of varied shape and of lower particle diameter than the original material (Figures 3 and 4).

**Figure 4.** Effect of pressure in the PGSS® processing of GMS particles: (**A**) GMS-1-67-120 and (**B**) GMS-1-67-200 particles.

Neurofuzzylogic software succeeded in modeling the influence of the parameters of pressure, temperature and nozzle diameter (inputs) on the output mean diameter (Table 2) with high predictability (R<sup>2</sup> > 90%) and accuracy (*p* < 0.01). The three parameters help to explain the variations in particle size, with temperature (submodel 1) having the main effect. An interaction between the pressure and the nozzle can be also observed (submodel 2).


**Table 2.** Inputs selected by FormRules ® for the di fferent outputs evaluated in this work, with their respective parameters to evaluate the quality of each model. The most relevant submodels are highlighted in bold.

The predictability is also reasonable for the percentage of fine particles (R<sup>2</sup> > 75%), a parameter indicative of process yield (Table 2). However, adequate accuracy was not achieved with such a small number of degrees of freedom. The model shows a main e ffect for the interaction pressure-nozzle, but temperature also a ffects process yield.

Variables studied do not explain su fficiently the variations in the standard deviation of the particle size distribution (R<sup>2</sup> < 75%). The particle size distributions with PGSS ® technique are broad and characterized by high standard deviations, probably higher than the variations promoted by the processing parameters (temperature, pressure and nozzle diameter) used in this research. Therefore, the ANN cannot define a good model for this standard deviation.

IF ... THEN rules, generated by the neurofuzzylogic software allows acquiring knowledge in an easy way (Figure A1). According to these rules, IF the temperature is low (up to 62 ◦C) THEN the mean particle size obtained is high (over 144.8 μm). The increase in temperature over 62 ◦C produces a decrease in particle size (Figure 3).

On the other hand, (IF) the pressure increase ( ... THEN) promotes a decrease in the particle size of the microparticles (Figure 4). This rule applies for both small and large nozzle diameters, being the variations in particle size wider when the large nozzle is used. Figure 5 represents the predicted results by the model for mean particle size for the large (Figure 5A) and small (Figure 5B) nozzle. This e ffect was related to the increased solubility of CO2 in molten GMS. At higher pressures CO2 solubility will increase and, upon depressurization, more nucleation bubbles will form due to CO2 supersaturation, breaking the lipid into smaller particles (Figure 4) [42,46]. Using the large nozzle diameter, pressure variations produced a more pronounced e ffect on the mean particle diameter.

**Figure 5.** Predicted results by the model for mean particle size for the ( **A**) large and (**B**) small nozzles.

Figure 6 shows the predicted values for the percentage of fine particles as a function of pressure and temperature. The increase in temperature leads to a reduction in the process yield, being especially important up to 62 ◦C. In the temperature range of the experimental design (57–67 ◦C), the Joule-Thomson coefficient is very similar for the 0–200 bar pressure range [47]. At higher temperatures, the positive Joule-Thomson effect contribution may not be enough to solidify the GMS when exiting the nozzle. Under these conditions, a significant fraction of GMS is in a semi-molten state when it reaches the precipitator and forms a crust in the walls of the vessel instead of forming SLMPs that deposit on the collector.

**Figure 6.** Influence of the parameters pressure and temperature on the yield of fine particle formation using: (**A**) the large nozzle diameter and (**B**) the smaller nozzle diameter.

The diameter of the nozzle also influenced the fine particle yield production. In general, the process performs better when using the small size nozzle. It has been reported that lower nozzle diameters led to smaller particle sizes for other lipid-based systems [48]. Differences in the effect of pressure were also detected depending on the size of the nozzle used. When a nozzle of smaller size is used, the increase in pressure causes a slight reduction in the percentage of fines obtained. This may be related to the production of even smaller particles that remain suspended in the CO2 and are therefore vented out. However, when the nozzle has a larger diameter, the effect is the opposite, and the process performance is improved with increasing pressure. The pressure drop of the lipid-CO2 melt through the nozzle is lower with larger nozzle diameters, leading to a decreased Joule-Thomson cooling effect. At higher pressures, this pressure drop effect is compensated by a higher CO2 content in the lipid melt and particles are able to solidify and reach the collector leading to higher fine particle yield [49].

The experimental values were compared with the values predicted with the model, showing high accuracy of the models for fine particle fraction (Figure 7A) and mean diameter (Figure 7B).

**Figure 7.** Parity plots of the predicted and experimental values of (**A**) mean particle size and (**B**) % of fine particles. Continuous diagonal line is a 45◦-slope line; dotted lines correspond to an envelope of tolerance of 10%.

### **3. Materials and Methods**
