Artificial Intelligence Deep Exploration of Influential Parameters on Physicochemical Properties of Curcumin-Loaded Electrospun Nanofibers

Citation for published version (APA): Khedri, M., Beheshtizadeh, N., Rostami, M., Sufali, A., Rezvantalab, S., Dahri, M., Maleki, R., Santos, H. A., & Shahbazi, M-A. (2022). Artificial Intelligence Deep Exploration of Influential Parameters on Physicochemical Properties of Curcumin-Loaded Electrospun Nanofibers. Advanced NanoBiomed Research, 2(6), Article 2100143. Advance online publication. https://doi.org/10.1002/anbr.202100143


Introduction
Curcumin (CUR), as a natural substance derived from the plant Curcuma longa, with its unique features as an antioxidant, [1] antitumor, [2] and anti-inflammatory agent, [3] has been widely used in drug delivery systems and tissue engineering applications. [4]UR can provide excellent free-radical scavenging activity, improving the neovascularization and increasing the migratory activity of various cells, including fibroblasts, dermal myofibroblasts, and macrophages onto the wound bed.The low bioavailability and solubility, low absorption, rapid metabolism, high sensitivity (after exposure to oxygen, light, moisture, and heat), and low in vivo stability of CUR resulted in the development of various formulations with different techniques, such as self-assembly, [5] spray drying, [6] emulsification, [7] solvent evaporation, [8] nanoprecipitation, [9] inclusion complexation, [10] liposome preparation, [11] freeze drying, [12] 3D printing, [13] Artificial intelligence (AI) methods are explosively considered in the design and optimization of drug discovery and delivery systems.Herein, machine learning methods are used for optimizing the production of curcumin (CUR)-loaded nanofibers.The required data are mined through the literature survey and two categories, including material-and machine-based parameters, are detected and studied as effective parameters on the final outcomes.AI results show that high-density polymers have a lower CUR release rate; however, with the increase in polymer density, CUR encapsulation efficiency (EE) increases in many types of polymers.The smallest diameter, highest EE, and highest drug release percentage are obtained at a molecular weight between 100 and 150 kDa and a CUR concentration of 10-15 wt%, with the polymer density in the range of 1.2-1.5 g mL À1 .Also, the optimal distance of %23 cm, the flow rate of 3.5-4.5 mL h À1 , and the voltage at the range of 12.5-15 kV result in the highest release rate, highest EE, and the lowest average diameter for fibers.These findings open up new roads for future design and production of drugloaded polymeric nanofibers with desirable properties and performances by AI methods.
electrospinning, [14] and electrospraying [15] to encapsulate CUR in different carriers for achieving better performance as compared with the free drug molecules. [16]he electrospinning process is a promising encapsulation method, which has been used to achieve submicrometer mats for hydrophobic bioactive agents, such as CUR.[19] Nanofibers fabricated by electrospinning have high porosity, an extremely large specific surface area, and proper pore interconnectivity. [20]Generally, the polymer fluid is pumped to a capillary spinneret by a feeding pump, and a high voltage electric field has been applied to the polymer solution.Increasing the voltage above the critical level leads to electrostatic charge repulsion and overcomes the solution surface tension, initiating the jet to go off from the top of the Taylor cone, resulting in the deposition of fibers on the collector. [18,21]ultiple parameters affect the features of the produced nanofibers and the final outcome, which could be divided into two categories, the machine-and material-based conditions.Machine conditions include voltage, flow rate, and needle distance from the collector, while material-associated parameters include concentration, viscosity, electrical conductivity, and solvent type. [17,19]In addition, the final features of the electrospinning process can be summarized in fibers' average diameter, drug release percentage (DR%), and encapsulation efficacy (EE) (Figure 1).Investigating the most influential input parameters on desired outputs via reviewing all in vitro and in vivo studies on CUR-loaded nanofibers is not possible because of the complexity of affecting multiple parameters simultaneously.Moreover, determining the optimized range of each influential parameter is not accessible through ordinary literature review and requires some robotic and information technologybased works.To achieve this point, using artificial intelligence (AI), and especially machine learning techniques, is valuable.
AI is a broad branch of computer science that focuses on organizing intelligent machines capable of accomplishing tasks, typically requiring human intelligence. [22]Therefore, AI is considered as a technology that allows machines to simulate human behavior.Moreover, machine learning is a subset of AI, allowing the machine to learn from past given data without explicit planning automatically. [23]Recently, AI has been used in medicine for drug design and discovery. [24]In an interesting AI study, Reker et al. [25] selected 100 self-assembling nanoparticles (NPs) among more than two million pairs to produce high drug loading and throughput NPs with more predictable properties.They claimed that using such methods can accelerate the progress of personalized drug delivery.Hathout et al. [26] reported a combination of multiple computational methods with AI in the prediction of drug loading in gelatin NPs.From another point of view, DR from 3D printed drug delivery systems has been investigated using various AI methods.Among the methods, artificial neural networks have been reported as the best approach for prediction and lowest error.Santana et al. [27] used information from Chemogenomic European Molecular Biology Laboratory (ChEMBL) Database and public sources in a Perturbation Theory Machine Learning (PTML) model to design NPs with optimal activity/toxicity profiles.
This article aims to develop an AI system based on all studies on CUR-loaded electrospun nanofibers to understand the most influential machine-based (voltage, flow rate, and distance between needle and collector) and/or material-based factors (polymer concentration, solvent conductivity, polymer density, polymer molecular weight (M w )) on the properties of the fibers, such as average diameter, CUR release percentage, and EE.Although some environmental parameters, such as temperature, humidity, and pressure, can also affect the procedure, they have been ignored in most studies, and utilizing them was unavailable.

Model Accuracies
The dataset of this study contains three outcome labels (average diameter, DR percentage, and EE), which for each six different models were tested.The accuracy and performance of each model was measured by their mean squared error (MSE), mean absolute error (MAE), root mean squared error (RMSE) values,R 2 scores for training data, and R 2 scores for testing data.In this part, model performance metrics (e.g., MSE, MAE, RMSE, and R 2 ) for each of the models and labels are provided and investigated.The reason for which all of these parameters are used together is that each of these parameters shows accuracy/ inaccuracy of a given model in a different way.The mathematical expressions for each of the aforementioned parameters are provided in Equations (1-5).
where y i , y, f ðx i Þ, x i , and n are the label value of i chosen sample in the dataset, the average label value in the dataset, the model predicted label value for ith sample in the dataset, the feature vector for sample ith in the dataset, and the number of samples in the dataset, respectively.Among these performance metrics, MAE and RMSE are in the same dimensions as the label value, while R 2 is dimensionless and MSE has a square dimension of the label values.The value of R 2 is between 0 and 1, with 1 showing the most accuracy and 0 showing no relationship.For more details, R 2 indicates the accuracy prediction.If R 2 value is 0, it indicates that the model is so weak, which is not able to predict the labels with enough accuracy.On the other hand, if the R 2 value is 1, it indicates that the model is strong in predicting the results, so that the results are accurate enough.RMSE, MSE, and MAE show how much model predictions have deviated from the dataset.Statistically, the RMSE shows if there is a large error (deviation) in even one prediction (this parameter is sensitive to even a few large errors), MAE shows that the overall deviation and is not sensitive to a few large errors.Table 1, 2, and 3 show performance metrics for different models for average diameter, DR%, and EE labels, respectively.Among different models, random forest tree, decision tree, and k-nearest neighbor models resulted in discrete stepwise values (instead of smooth trends), and despite their quite good performance metrics, they were put aside.The neural network model also experienced severe overfitting due to the low number of data samples and model complexity and it was put aside too.The linear model was so simple model and resulted in underfitting due to its simplicity.Finally, the support vector machine model with the radial basis function kernel was chosen as the best model due to its balanced complexity and no underfitting/overfitting results.

Polymer Properties
The polymer molecular weight (M w ) is an important criterion in the processing and application of nanofibers. [28,29]The size of the polymer chain is determined by the polymer M w , which affects the viscosity of the solution at a given concentration.As a result, viscosity is in close contact with polymer molecular weight and density.Typically, a polymer with high M w produces higherviscosity polymer solution at an identical concentration than a lower one, which causes a higher chain entanglement that is necessary to form continuous jet and uniform nanofibers. [30]These polymer chain entanglements control the stability of the jet and inhibit the formation of beads and droplets.Possessing an appropriate viscosity is one of the essentials for electrospinning to form nanofibers. [31] In general, the diameter of nanofibers decreases with the increase in M w, [29] and it has been shown that high M w is preferable for the electrospinning process and polymers with lower M w tend to form beads rather than fibers. [32]owever, an optimal range of M w can be determined before expensive and laborious experiments.In this regard, we evaluated the effect of polymer M w on the fibers' average diameter and their features on CUR-loaded nanofibers (Figure 2A).We found that with increasing the M w of polymers from 1 to 150 kDa in various solutions, the average diameter of nanofibers decreases significantly.Also, at the M w of 150 kDa, the smallest diameter of nanofibers is produced and above this value, fibers' diameter increased.There is no continuity in jet producing and forming droplets, and bead formation is more possible due to low chain entanglement when low-M w polymers are used.In contrast, in higher-M w conditions, a uniform nanofiber with a small diameter can be formed due to sufficient chain entanglement and the effect on viscosity.Furthermore, M w is an effective parameter on the EE that has been clearly verified by EE of CUR-loaded electrospun nanofibers.The deconvolution of the reported EEs in the mined data allowed us to distinguish that the highest loading in nanofibers is in small sizes (<400 nm) and the most uniform nanofibers.In fact, the highest EE can be achieved approximately at M w of 150 kDa, whereas above this value, EE of the CUR drops.DR% was also assessed for the CUR-loaded nanofibers.DR is plotted versus M w and the highest DR rate is found for nanofibers with the highest EE.The highest DR% was also observed in polymers with a M w range of 100-150 kDa, which had the smallest diameter and the highest EE.At higher M w , the release rate of CUR is reduced as the loading at these molecular weights was lower.Overall, it can be understood that the smallest diameter, highest EE, and highest DR% can be obtained at M w between 100 and 150 kDa that is highlighted in the plot shown in Figure 2A.
We also investigated the impact of polymer density on the CUR-loaded nanofibers' properties and performance.Figure 2B represents the density influence on the diameter of the produced nanofibers that notes that increasing the polymer density from 1.2 to 1.5 g mL À1 results in the smaller diameter of nanofibers.The lowest diameter of nanofibers is obtained in polymers with a density of 1.5 g mL À1 , while fibers' diameter increased significantly in higher density values.The density of the polymer affects the viscosity of the solution, and with growing polymer density, viscosity and molecular chain entanglement are increased.This reduces the fiber thickness and produces uniform fibers at a critical polymer density.Hence, the range of 1.2-1.5 g mL À1 polymer density was assigned as the optimal range of polymer density for producing CUR-loaded nanofibers.Furthermore, it can be seen that the highest EE (%) of CUR was observed at the polymer density range of 1.6-1.8g mL À1 , and it is possible that the uniform fibers were produced in this range.Based on these results, high-density polymers have shown lower CUR release rate; however, with increasing polymer density, CUR trapping in the polymer may be increased.

Concentration of the Polymer Solution
Generally, it has been reported that electrospun nanofibers are strongly influenced by the concentration of polymer and drug solutions. [33]The polymer concentration strongly affects the solution viscosity and subsequently the surface tension of the polymer solution.The elongating of the charged jet is affected by changing the concentration, viscosity, and surface tension of the polymeric solution.With inappropriate concentrations of solution, beads with different sizes and amount can be formed on the nanofibers. [34]If the concentration of the polymer solution is low, the viscosity is low, in which case the electric field and the surface tension of the solution cause the charged jet to break before reaching the collector and results in bead formation. [35]ith a slight increase in concentration, a mixture of fibers and beads can be formed.Increasing the concentration of the polymeric solution will lead to enhancing the chain entanglement among the polymer chains with an increase in viscosity.These chain entanglements overcome the surface tension and finally result in preparing the uniform bead-free electrospun nanofibers.The concentration at which the most uniform, nonbeady nanofibers are formed is the critical concentration of the polymeric solution.Above the critical concentration, by increasing concentration and viscosity, the flow of polymer solution and jet formation stops, in which case the polymer solution dries at the tip of the needle and prevents the flow of solution.Eventually, it leads to a production defect or an increase in the size of the beads. [36]erein, we also investigated the influence of concentration on CUR-loaded nanofibers.By monitoring the effect of polymer concentration on the average diameter distribution of CURloaded nanofibers, we found that concentration enhancement from 1% to 10% for different polymers would increase the average diameter of nanofibers (Figure 3A).It has been shown that the largest average diameter of nanofibers was gained at a polymer concentration of 10% in various polymeric solutions and enhancing the concentration of polymers from 10% to 30% significantly decreased the diameter of fibers.We observed that the lowest fibers' average diameter was created at a concentration of 30-35% for all polymers in the present study, and uniform fibers could be obtained in this area.
Hence, in this study, a 30-35% polymer concentration is considered a critical optimized concentration (highlighted in Figure 3A).By evaluating the EE (%) of CUR in various studies, we observed that the maximum EE (%) was obtained at the polymer solution concentration of 15%, whereas this concentration resulted in high fiber average diameter, which is inappropriate.Also, regarding the effect of polymer solution concentration on the DR (%) of CUR, we found that the high CUR release was obtained when utilizing the concentration of polymer solution below 20% and/or above 40%, as these ranges possess the high EE (%).
Investigating the results of CUR encapsulation in electrospun nanofibers showed that increasing the CUR concentrations to 1-10% resulted in significantly decreased diameter of the fibers (Figure 3B).The lowest average diameter of the nanofibers was obtained at a CUR concentration of 12%, while higher concentrations led to an increase in fiber diameter.The results obtained from AI demonstrated that a concentration of 10-15 wt.-% is the optimal CUR concentration to prepare mats possessing the smallest diameter and most uniform nanofibers.This can also affect the morphology of nanofibers.For instance, Merrell et al. [37] reported that with 15-20% concentration of CUR, nanofibers can reach beadles morphology.Also, by increasing the loading concentration to 1-20 wt%, the EE (%) of CUR increased and the highest EE (%) was reached at a concentration of 20 wt%.As the smallest fiber diameter was obtained at a concentration of about 15 wt%, we noted that the most uniform fibers are produced at this CUR concentration, possessing a high amount of EE (%).In addition, an increase in the CUR concentration up to 20 wt% resulted in DR% to surge to an almost constant level with a slight slope.

Solvent Electrical Conductivity and Release Time
The electrical conductivity of the solution plays an essential role on the tensile strength of the polymer droplet and the formation of the Taylor cone and the jet. [38]This is achievable through the electrical repulsion of polymer solution.Solvents with various electrical conductivity can affect the final electrical conductivity of the polymer solution. [39]In a solution with low electrical conductivity, the surface of the droplet would not have any charge to make a Taylor cone; hence, electrospun nanofibers cannot be formed. [40]However, with increasing electrical conductivity of the polymeric solution to a critical value, electrospun nanofibers would be prepared, and also it might cause a considerable decrease in the fiber average diameter. [41]The Coulomb force among the electrical charges on the fluid surface and the force due to the external electric field is essential for continuing the electrospinning process.The formation of the Taylor cone is mainly dependent on the electrostatic force of the surface charges created by the applied external electric field.However, a polymeric solution will have enough free charges to move onto the surface of the polymeric solution, make a Taylor cone, and initiate the formation of electrospun nanofibers. [41,42]s the electrical conductivity of the polymer solution was not available in most studies, the electrical conductivity effect of solvents used in preparing CUR-loaded nanofibers was investigated.It is observed that with increasing the solvent electrical conductivity, the average diameter of the produced nanofibers decreases slightly (Figure 4A).Apparently, higher electrical conductivity produces smaller nanofibers.However, it is evident that solvents have been selected for the electrospinning process that renders the spinnability to polymer solution; hence, the critical point for nanofiber formation is not clear from this diagram.The influence of solvent electrical conductivity on the EE (%) of nanofibers has shown that by increasing the electrical conductivity up to 5500 μS cm À1 , EE increases to 80%.The diameter of the produced nanofibers was reduced, and most uniform nanofibers were obtained in highly electrically conductive solvents.In addition, the highest DR was achieved at high electrical conductivity, in the vicinity of 5000 μS cm À1 .
One of the tests performed on CUR-loaded nanofibers is the DR test based on the duration in the release medium.The effect of DR time on the DR rate in produced nanofibers is demonstrated in Figure 4B using AI.It is observed that in nanofibers containing CUR, the release rate increases over time.Therefore, a continuous release occurred from the first day to the 25th day.The maximum release rate was on days 20-25, and after this day, the release rate decreased.

Machine-Based Influential Parameters
As mentioned before, the second category of the effective parameters on the morphology and performance of nanofibers is the machine-based variables, including solution flow rate, the distance (between collector and the needle), and applied voltage.During the production process, the solution is pumped from the syringe pump, and the nanofiber is produced by a polymer drop at the tip of the metallic needle. [21]The flow rate of the polymer solution affects the preparation of uniform nanofibers. [43]tudies show that there is a critical flow rate for the formation of uniform nanofibers and that for each polymeric solution, it should be determined, as it varies for different polymers. [21]bove or below this critical flow rate, the fibers are in a bead-like manner.By increasing the flow rate above the critical point, the diameter of the nanofibers increases due to the low tensile forces and incomplete drying of the jet during the flight between the metallic needle tip and collector.In addition, the bead shape in the nanofibers increases in this condition. [39,44]Figure 5A shows that the minimum diameter of CUR-loaded nanofibers is at the flow rate of 1 mL h À1 , and with increasing the flow rate above 1 mL h À1 , the diameter of produced nanofibers increases continuously and can reach to the microscale region.
According to the AI analysis performed on the studies, a flow rate of 1 mL h À1 as a critical flow rate is justified for producing CUR-loaded nanofibers.Evaluating the flow rate effect on the EE of CUR-loaded nanofibers demonstrates that the highest distribution of CUR loading was gained at a flow rate of 4 mL h À1 .As shown, in the flow rate zone of 4 mL h À1 , the diameter distribution of nanofibers is in the range between 2700 and 2800 nm.It might be the highest uniformity of produced nanofibers in which the EE (%) is higher than other flow rates.Furthermore, the highest CUR release rate could be obtained through the highest-loaded CUR.We observed that in the region of 3.5-4.5 mL h À1 flow rate, where the highest encapsulation of CUR was achieved, the highest release rate also occurred.
The distance between the collector and metallic needle tip plays an essential role in electrospun nanofibers' average diameter and morphology.The distance affects the average diameter and nanofiber morphology as it depends on the evaporation rate, deposition time, and instability interval or whipping. [45]If the distance between the needle tip and the collector is short, the solvent cannot evaporate entirely before reaching the collector.As a result, fibers with a large diameter are produced.It has been shown that beads and bead fibers were formed at large distances between the tip of the needle and the collector. [45,46]Therefore, an equilibrium or critical distance is required to produce uniform, bead-free nanofibers with a nanodiameter distribution. [47]tudies have used a distance of 7.5-30 cm between the needle and the collector for producing CUR-loaded nanofibers through the electrospinning process.AI analysis revealed that the optimal distance for this procedure with the lowest fibers' average diameter is %23 cm (Figure 5B).It is also observed that the largest average diameter of the nanofibers occurred at the distance of 13 and 30 cm between the needle and the collector.
Regarding the effect of the distance between the needle and the collector on the EE (%), Figure 5B shows that the EE range of 73-93% was achieved in a distance of 7.5-30 cm in several studies.We observed that the lowest EE (%) was obtained for nanofibers produced at a distance of 17.5 and 20 cm between the needle and collector.This indicates that higher EE (%) was obtained at quite long or short distances.Also, according to the AI analysis, the needle and the collector distance affect the release percentage of CUR, while the highest release was occurred for distances of 25-28 cm, where the fibers had the highest loading of CUR.
Generally, a high-voltage power supply into a solution via a stainless needle causes a shape change from a spherical droplet to a Taylor cone.Studies have shown that forming ultrafine micro-and/or nanofibers occurred at a critical voltage. [17,48]en a high voltage electric field is applied to the polymer droplets, the droplet surface is charged, and the electrostatic force is overcome on the droplet surface tension, and the electric jet forms.Forming various shapes and smaller-diameter nanofibers positively correlates with an increase in the applied voltage, which is attributed to the stretching of the polymer solution in correlation with charge repulsion within the polymer jet. [17,49]AI analysis data revealed that by enhancing the voltage above 30 kV, a progressive increase in the nanofibers' average diameter was obtained (Figure 5C).The raise in the average diameter and formation of beads with an increase in the applied voltage are attributed to the decrease in the size of the Taylor cone and increase in the jet velocity for the same flow rate.
It has been shown that the highest EE (%) occurs at the range of 12.5-15 kV.At this voltage, the most uniform and ultrafine nanofibers of various polymers were produced.As a result of the nonuniformity of nanofibers and multiple beads or bead-fiber forming, the EE (%) decreases at very high voltages.It is also known that the highest release of CUR occurs when the highest amount of this substance is loaded, normally, at a voltage of 15 kV, which has the highest EE (%), with a favorable release rate of 62%.
Figure 5D represents the Spearman's correlation partial heatmap for the detected effective factors during the study.The map shows a consistent correlation between the factors.Clearly, CUR

Conclusion
The electrospinning technique was used as a suitable process to produce CUR-loaded nanofibers for various biomedical applications.In the current study, using data mining from literature, the most prominent parameters on producing nanofibers and subsequently the properties of produced CUR-loaded nanofibers were investigated through AI methods.The determinant parameters are categorized into two groups, including material-and machine-based factors.The possible range of individual factors and their impact were evaluated, as well as the optimal range for each factor.Through the AI method, all the CUR-loaded nanofibers' production parameters were optimized and also forecasted.
Based on the results obtained by this AI study, M w between 100 and 150 kDa, CUR concentration of 10-15 wt%, with the polymer density in the range of 1.2-1.5 g mL À1 as optimized material-based parameters, along with the distance of %23 cm, the flow rate of 3.5-4.5 mL h À1 , and the voltage of 12.5-15 kV as optimized machine-based parameters, are proposed for future studies.Performing AI study, the need for repeating various polymers, solvents, and concentrations decreased, while the possibility of predicting the output results of fiber average diameter, EE, and DR rate increased for various parameters' conditions.

Experimental Section
This study was performed in four stages, including data preprocessing, feature importance estimation, model training and testing, and optimization.Table S1, Supporting Information, gathers all the references that we have taken for this study.The dataset in which this study was performed consisted of nine features and three labels.Dataset features were polymer molecular weight, polymer density, polymer concentration, CUR concentration, solvent electrical conductivity, flow rate, distance, voltage and DR time, while the data labels were average diameter, DR%, and EE (%).
In the data preprocessing stage, at first, the data samples that were missing average diameter values were removed.Then, the missing features in each data sample were approximated using a fast k-nearest neighbor imputation algorithm with 30 neighbors (data samples which had minimum Euclidician distance to the original point in n-dimensional space).The imputation process was performed using imyute library.After data imputation and approximating missing features in data samples, each feature/label was scaled to [0, 1] interval using min-max scaling.In this way, the model training procedure was more robust and efficient.The min-max scaling for a feature/label Z of a data sample is shown in Equation (6).
where Z is the deired feature or label to be scaled and Z Min , Z Max , and Z Scaled are minimum, maximum, and scaled values of desired feature or label, respectively.The feature importance stage was performed to understand the efficacy and the degree of importance for each of the data features for a given label.To measure feature importance for a given label, spearman's correlation coefficient was calculated between that specific label and different features.The spearman's correlation coefficient was chosen over other methods of feature importance measurements due to its computational efficiency, nonlinearity support, and simplicity.The spearman's correlation coefficient for pair feature/label of X/Y is indicated in Equation ( 7) and (8).
where RðX i Þ and RðY i Þ are rank values of ith data in X/Y feature/label.In the model training and testing stage, six different models were trained and tested for different labels.Random forest tree, decision tree, linear regression, support vector machine, k-nearest neighbor, and neural network models were the models which were trained and tested.All of the aforementioned models, except the neural network model, were built, trained, and tested using the scikit-learn library.The neural network model was built, trained, and tested using the tensorflow and keras libraries.Before model training, 80% of the whole dataset was picked to be used as training set and the remaining 20% was picked to be used as testing data.To check if a model was overfitting, R 2 scores of training and testing data were compared and if there was a big difference between these two scores, the model was overfitting.After performing model training, testing, and selecting the suitable model to predict each of the labels, these models were used as descriptor functions and based on their predictions, and the optimal features were obtained in an optimization process.The optimization process was performed using particle swarm optimization (PSO) algorithm as it was an evolutionary algorithm being able to find more robust results.
To make the PSO algorithm more robust and efficient, its parameters were chosen from good parameters for PSO (GPPSO), which are shown in Table 4.
In Table 4, ω, ϕ P , and ϕ G refer to inertia weight, cognitive coefficient, and social coefficient, respectively.
Statistical Analysis: All data were gathered from published experimentally original studies (Table S1, Supporting Information).Each data was presented as a single value without AE standard deviation (SD).For this purpose, the average data was taken wherever necessary.Sample sizes (n) for each study were similar to the corresponding reports.All figures and statistical analyses were performed via Origin Pro (2019b Build 9.6.5.169).Table 5 shows the overall statistics of the dataset, which contained 157 data samples.

Figure 1 .
Figure 1.Schematic of the workflow in the study to determine the effective parameters in the preparation of nanofibers regarding inputs and outputs in the machine learning system.The figure was created by PowerPoint.

Figure 2 .
Figure 2. Effect of polymer A) M w and B) density on the size, EE, and DR% of CUR-loaded electrospun nanofibers.

Figure 3 .
Figure 3.The effect of A) polymer and B) CUR concentration on CUR-loaded electrospun nanofibers.

Figure 4 .
Figure 4.The effect of solvent electrical conductivity and DR time on CUR-loaded electrospun nanofibers.

Figure 5 .
Figure 5.The effect of machine-based parameters, including A) flowrate B) voltage, and C) distance on CUR-loaded electrospun nanofibers.D) Spearman's correlation coefficient between different features/labels.

Table 1 .
Average diameter trained model performances.

Table 4 .
Parameters used in this work based on GPPSO.

Table 5 .
The overall statistics of the utilized dataset.