Aleatico grape bunches ( Vitis vinifera L.) were manually harvested (240 ± 3 g/L of sugar) in 2022 at Famiglia Cotarella-Falesco in Montecchio (TR), Italy and then were placed in perforated crates (about 6 kg grape each crate).

The grapes (6 crates for each sample) were subjected to dehydration: i) at room temperature (RT) of 21 ± 2 °C and 45 ± 5 % RH (relative humidity) with an airflow of 1 m/s, or ii) in a cold room (CR) at 9 ± 2 °C, 60 ± 5 % RH (maintained using a dehumidifier) with an airflow of 1 m/s. Temperature and RH sensors were used for daily monitoring developed by DIBAF (University of Tuscia, Viterbo) ( Santini et al., 2023); airflow was measured by using an anemometer (Lutron LM-8000, Coopersburg, PA, USA). The 6 crates were weighed each day and dehydration was stopped at approximately 35 % weight loss when the grape reached about 400 g/L of total sugars. Dehydrated grapes from RT or CR were poured into a crusher–destemmer and each must divided into two lots: the first lot was immediately fermented to produce sweet wine (Sweet RT and Sweet CR); The other was used to produce fortified wines (Fortified RT and Fortified CR).

To produce fortified wines from RT or CR, each must was roughly filtered (metal net) to eliminate debris and divided into three 5 L glass jars for each sample. Each jar had 95 % (v/v) of alcohol (food grade) added until an alcohol content of 15 % (v/v) was reached; after one month, the wines were bottled in 0.75 L glass bottles with a silicon stopper and stored at 15 °C.

To produce passito sweet wines, must from dehydrated grapes at RT or CR, without filtration, were divided into three 5 glass jars, and 8 g/hL of K ₂S ₂O ₅ and 15 g/hL of Zymaflore X5 yeast (Laffort srl, Tortona (AL), Italy) added to each one. At day 3 from the start of fermentation, 20 g/hL of fermentation activator (Nutriferm® Vit Flo, Enartis srl, San Martino (NO), Italy) were added. Pigeage was performed every day. The fermentation process was blocked adding 12 g/hL of K ₂S ₂O _5, at a must sugar concentration of around 160–180 g/L. Then, the four wines (Fortified RT, Fortified CR, Sweet RT and Sweet CR) were bottled in 0.75 L glass bottles with silicon stopper and stored at 15 °C. No malolactic fermentation was performed.

Grapes were analysed at harvest and the end of the dehydration process (3 samples of 50 berries from different crates). Wine analysis was performed on finished wines on 3 bottles for each sample. Chemical determinations were carried out by a calibrated Fourier transform infrared WineScan™ FT 120 (Foss Analytics, Hillerod, Denmark) to determine: sugars (g/L hexoses), pH, titratable acidity (tartaric acid g/L), volatile acidity (g/L acetic acid), malic acid (g/L), tartaric acid (g/L), citric acid (g/L), gluconic acid (g/L), ash (g/L), YAN (g/L), total extract (g/L), glycerol (g/L), total anthocyanins (mg/L malvidin) and total polyphenols (mg/L gallic acid). The accuracy of the WineScan™ analyses was validated, periodically, by destructive analyses performed using the reference OIV methods as reported by Bianchi et al. (2021).

The chromatic characteristics were detected by a Benchtop CLM-196 colorimeter (Eoptis-38121 Trento (TN)-ITALY). The colour values were expressed using the native coordinates CIE L*a*b* and the cylindrical coordinates, h* and C*, calculated according to Modesti et al. (2021). The colour differences among samples (∆E* _ab) were also calculated as previously reported ( Bianchi et al., 2025a) and are expressed in CIELAB units.

Volatile organic compound (VOC) analyses were carried out as previously reported ( Mastrangelo et al., 2023). In particular, 10 mL of wine sample and 100 μL of a 2-octanol solution at 500 mg/L was added as an internal standard. The sample was deposited on a Hypersep Retain Pep (Thermo Fisher Scientific, Milan, Italy) cartridge (60 mg), activated with 2 mL dichloromethane, 2 mL methanol and 2 mL water. The analytes were eluted with 5 mL of dichloromethane, collected in sovirel on the bottom of which 2-grams of anhydrous sodium sulfate had been inserted and placed in the freezer overnight. Prior to GC-MS analysis, samples were filtered with a cellulose filter to remove sodium sulfate and concentrated to a final volume of 200 μL under a stream of N ₂.

The GC apparatus consisted of a Trace GC ultra-gas chromatograph with a Trace DSQ with quadrupole mass detector (Thermo Fisher Scientific, Milan, Italy) and a Stabilwax DA capillary column (Restek, Bellefonte, PA, USA; 30 m, 0.25 mm i.d., and 0.25 μm film thickness). The carrier gas was He with a constant flow of 1.0 mL/min.

GC temperature ramp was programmed as follows: from 45 °C (maintained for 1-minute) to 100 °C (maintained for 1-minute) at 3 °C/min, then to 240 °C (maintained for 10 minutes) at 5 °C/min. The injection was performed at 250 °C in splitless mode and the volume injected is 1 µL.

The identification of compounds was carried out as previously reported ( Mastrangelo et al., 2023), following a triple criterion: (i) by comparing compound mass spectra and retention time with those of pure standards, (ii) matching their respective mass spectra with those present in online libraries Willey and NIST 08, (iii) by comparing linear retention index (LRI) calculated under our analytical conditions with already published LRI calculated on polar columns.

Quantification was carried out by normalisation of integrated areas in total ion current with the area of 2-octanol (internal standard). Analyses were performed in triplicate and GC-MS parameters were obtained by using Xcalibur software (version 4.1, Thermo Fisher Scientific, Milan, Italy).

The comprehensive analysis of polyphenols within grape and wine samples involved high-performance liquid chromatography (HPLC) using the method previously reported ( Tarì et al., 2024).

1 mL of grape juice or wine, diluted at a 1:1 ratio with phase A, was extracted from the samples. The obtained sample was filtered through a 0.45 μm PVDF (Polyvinylidene fluoride) filter before being injected into the HPLC system ( Tarì et al., 2024): PU-2089 Plus quaternary pump (Jasco International Co., Ltd., Tokyo, Japan), equipped with a degasser, an AS-2057 Plus autosampler (Jasco International Co., Ltd., Tokyo, Japan), and a CO-2060 Plus column oven (Jasco International Co., Ltd., Tokyo, Japan). Detection was conducted using a UV-2070 Plus visible detector (Jasco International Co., Ltd., Tokyo, Japan). The acquired data underwent processing via ChromNAV (software version 2.3). The separation took place using a DionexAcclaim ^®120 C18 column, 5 μm, 4.6 × 250 mm (Thermo Fisher Scientific, Milan, Italy), maintained at a constant temperature of 40 °C, with a mobile phase flow rate set at 0.5 mL/min. The mobile phase comprised a ternary gradient: solvent A = 50 mM ammonium dihydrogen phosphate adjusted to pH 2.6 with phosphoric acid, solvent B = 20 % solvent A and 80 % acetonitrile, and solvent C = 0.2 M orthophosphoric acid adjusted to pH 1.5 with NaOH and the solvent gradient used as reported in Pettinelli et al. (2024).

For the qualitative and quantitative analysis of individual polyphenols in the studied matrix, the construction of calibration curves for phenolic compounds of primary interest was performed. Given the extensive variety of polyphenols present in grape and wine matrices, a total of 33 standards (Sigma-Aldrich, Steinheim, Germany) were thoughtfully selected for this purpose.

The electronic nose (E-nose) employed in this study was designed and assembled at the University of Rome Tor Vergata, incorporating eight quartz microbalances (QMB). The functionalisation scheme of the metalloporphyrin defining the pattern of the eight sensors is as follows: QMB1: Mn-TPP (tetraporphinin), QMB2: Co-p-OCH3-TPP, QMB3: Sn-TPP, QMB4: Rh-TPP, QMB5: Co-p-TPP, QMB6: Cr-TPP, QMB7: Co-TPP, QMB8: Ru-TPP ( Mastrangelo et al., 2023; Saevels et al., 2004). These sensitive molecules, synthesised at the Department of Chemical Sciences and Technologies of the University of Rome Tor Vergata, were specifically chosen for their proven sensitivity to VOCs. In these sensors, minute changes in mass (Δm) on the absorbing layer of the quartz surface induce a frequency shift (Δf) in the electrical output signal of the oscillator circuit. Within a range of minor alterations, the change in frequency (Δf) is directly proportional to the change in mass (Δm). The variation in the oscillation frequency of quartz microbalances occurs upon the adsorption of volatile molecules from the sample's gaseous phase. Each sensor is linked to a 3-point oscillator, and the output signal relies on the resonant frequency of the microbalance. The instrument is powered through a micro USB connector, also serving for data transmission. Frequency measurement is executed by a calibrated counter with a reference resonator, boasting a frequency resolution on the order of 0.1 Hz. The mass measurement resolution is 0.17 ng. Before commencing the analysis, the initial step involved a nose-cleaning phase by connecting the tubes to a filter containing calcium chloride for over an hour until stability signs were evident on the scales. The analyses were conducted on 10 mL of wine, incubated in closed 25 mL vials (fitted with silicone septum) at 30 °C for 20-minutes. The equilibrated headspace was then extracted for 20-minutes using a stream of filtered air and directed into the electronic nose sensor cell. Each sampling unit comprised 3 measurements. The assessment of aromatic enrichment in the headspaces involved evaluating the signal of the oscillation frequency of individual sensors. This signal was compared to a null measurement reference, which consisted of air measurements devoid of odorous molecules and humidity. The latter served as a flow of pure air, cleaning the electronic nose for 20-minutes and establishing the reference signal. The ensemble of sensor signals formed distinct models (fingerprints) encoding the global composition of the headspace.

The objective of the analysis using the NIR-AOTF spectrophotometer (Luminar 5030 Miniature Hand-held AOTF-NIR Analyzer, Brimrose, Baltimore, MD) was to assess the discriminative capacity of the acquired spectra in identifying significant differences among wine samples ( Antolini et al., 2021). The operational procedure involved placing the wine inside a test tube, directly in contact with the external probe of the instrument. Measurements were conducted in specular reflectance (R), and the "raw" spectra were acquired in transmittance (T). The spectral acquisition spanned the range of 1100–2600 nm, with a measurement step of 2 nm, and an acquisition speed of 10 scans/s was employed. A total of 20 spectra were acquired. Subsequently, the measured spectra were transformed into absorbance values (log 1/T). Each sample underwent the acquisition of 9 spectra, which were then averaged to obtain 3 representative spectra for each sample thesis.

Furthermore, the working spectral range (1100–2600 nm) encompasses key regions associated with molecular vibrations and electronic transitions, giving information about the composition and characteristics of the wine samples. The adoption of 10 scans/s for acquisition ensures a high level of data precision, contributing to the reliability of the analysis ( Dambergs et al., 2015; Páscoa et al., 2020). The conversion of spectra into absorbance values enhances the interpretability of the data, facilitating the identification of distinctive features and patterns within the samples.

The data analyses were conducted using the JMP Pro 17 software package from SAS Institute located in Cary, NC, USA. In assessing the statistical significance of the experimental data, each sample underwent triplicate analysis. The determination of differences among samples was performed through One Way Completely Randomized ANOVA using CoStat, a software tool from Cohort (version 6.451, Pacific Grove, CA, USA), coupled with Tukey’s HSD multiple mean comparison test ( p ≤ 0.05).

The numerical data of NIR and E-nose, once the acquisitions were completed, were assembled within a matrix for subsequent multivariate statistical reprocessing to obtain a discriminative evaluation of non-qualitative supervision (pattern recognition). This approach was pursued on pre-treated data with a normalisation filter (autoscaling) and applying partial least square discriminant analysis (PLS-DA) using leave one out as a validation method.

These analyses and graphical representations were conducted using Matlab R2021a software (Mathworks), supplemented by the PLS Toolbox tool from Eigenvector Research.

A significant difference in dehydration time was observed between RT and CR (Figure 1). At RT, a sugar concentration of approximately 400 g/L was reached in 19 days, whereas the same concentration required 47 days in CR (Table 1). Despite the different dehydration times, both methods resulted in a weight loss of about 40 %. Moreover, the R² values of the regression lines were very high, indicating a strong correlation.

Figure 1. Regression line of the weight loss (%) during the dehydration period (days). Data are the mean of 3 crates (Bars indicate the standard deviation). RT = Room temperature; CR = Cold room.

At the end of the dehydration process, as expected, a significant increase in sugars and total extract compared to the initial sample was observed, without a significant difference between the two dehydrated samples (Table 1). Increases in pH and titratable acidity were also observed, especially in the CR sample. This is a typical behaviour of grape dehydration where pH and titratable acidity rise together. This is due to the buffer effect as a consequence of cell wall disruption with the release of cations that produce salts with weak acids such as tartaric and malic acids ( Bellincontro et al., 2016; Zoccatelli et al., 2013).

Volatile acidity increased significantly due to the physiological formation of acetic acid as well as the increase of glycerol due to the glyceropyruvic fermentation in the berry under water stress conditions ( Bellincontro et al., 2016; Pettinelli et al., 2022). Similarly, gluconic acid increased due to glucose oxidation. Differences were observed for total anthocyanins as well as a significant increase of polyphenols was observed in dehydrated grapes, regardless of the condition (Table 1). This increase can be probably attributable to the concentration effect but also the triggering of the water stress on the phenylpropanoid pathways ( Bellincontro et al., 2016; Centioni et al., 2014; Zoccatelli et al., 2013).

Table 1. Chemical parameters of Aleatico grapes at harvest and the end of the dehydration process in the RT and CR conditions.

Data are the means (± standard deviation) of 3 samples of 50 berries each from different crates using 3 WineScan ^TM analyses. Different letters in each row indicate statistical differences ( p ≤ 0.05). n.d. = not detected.

After dehydration, a concentration effect was observed for most of the individual phenols identified (Table 2). Comparing a fast dehydration rate at room temperature (RT) with a slow dehydration rate at cold temperature (CR), a significantly higher concentration of anthocyanins, flavonols, flavanols, and two hydroxybenzoic acids is evident in CR grapes. By calculating the 40 % weight loss based on the initial grape weight to roughly estimate the concentration effect, different trends are observed for anthocyanins. In CR, the final anthocyanin concentration corresponds to the expected value, suggesting a concentration effect due to water loss. Conversely, in RT, the final anthocyanin concentration is lower than expected, indicating a loss of anthocyanins during dehydration rather than a mere concentration effect. In the other classes, flavonols were concentrated in both samples, while flavanols and the two hydroxybenzoic acids showed values significantly higher than expected, particularly in the CR sample. This result confirms the anabolic effect already reported in the literature ( Bellincontro & Mencarelli, 2022). In the case of anthocyanins, only a concentration effect was observed in the CR sample (not synthesis, as there was an increase in flavonols, and it is known that the synthesis of these compounds is competitive). In contrast, in the RT sample, the loss of anthocyanins likely occurred due to an oxidation process at room temperature, as anthocyanins are sensitive to oxidation ( Bianchi et al., 2022; De Rosso et al., 2010; Figueiredo-González et al., 2013b; Toffali et al., 2011).

Table 2. Single phenolic compounds (mg/L) in Aleatico grapes at harvest and the end of the dehydration process in the RT and CR conditions.

Data are the means (± standard deviation) of 3 samples of 50 berries each from different crates. Different letters in each row indicate statistical differences ( p ≤ 0.05). n.d. = not detected.

In Table 3, the main oenological parameters are detailed with their respective significant differences for the wines produced by the four types of grapes.

One of the most obvious results is a higher alcohol concentration in Fortified wines, a result of the addition of ethanol as well as a higher concentration of sugars, as the alcoholic fermentation process was not carried out.

The higher pH (and the lower titratable acidity) in Fortified wines can be explained by the alcoholisation process without fermentation. It is well known ( Chidi et al., 2018; Thoukis et al., 1965) that fermentation acidifies the wine because of the formation of non-volatile acids. As expected, alcoholic fermentation favoured a greater concentration of volatile acidity, due to the high sugar concentration of grapes as well as yeast metabolism. The Fortified wines cannot exhibit this increase, as the addition of alcohol has a sterilising effect, preventing further changes. Additionally, there is no physiological formation of compounds in yeast during fermentation that would contribute to this increase. Among acids, the citric acid showed a significantly higher concentration (4–5 folds) in Sweet wines (Table 3). This behaviour could indicate that during fermentation, the yeasts begin the Krebs cycle thanks to a higher concentration of acetic acid which provides the acetyl group for the formation of citric acid, which probably accumulates because of the difficulty of the Krebs cycle in moving forward ( Mendes Ferreira & Mendes-Faia, 2020). Indeed aconitase, the enzyme which transforms citrate into isocitrate, is very sensitive to oxidation; in mammalian cells, it is the enzyme in the Krebs cycle that decreases during aging ( Yarian et al., 2006). Grape cells under the dehydration process not only are water stressed but also senesced ( D’Onofrio et al., 2019; Zenoni et al., 2016).

The total extract was significantly higher in the Fortified wines due to the higher sugar concentration. The anthocyanins and polyphenols had a significantly lower concentration in the Fortified wines (Table 3). The initial formation of anthocyanin−pyruvic acid adducts has been seen concurrent with the degradation of anthocyanidin monoglucosides in ageing Port ( Mateus & de Freitas, 2001), but Ossola et al. (2017) found that Fortified wines from dehydrated grapes were chromatically unsatisfactory due to the low content of total anthocyanins. Anyway, the fortification does not permit the extraction of chemical compounds which occurs during fermentation. Thus, the effect of postharvest water stress on grape berries is very critical for the colour of Fortified wines, much more than in Sweet wines (Table 3). The CIELAB colour data confirm that sweet wines have lower lightness (L*) values, indicating a significantly reduced white component and, consequently, a darker colour than fortified wines. This may be attributed to the dilution effect caused by the addition of alcohol, as observed for all the analysed chemical components. Regarding the a* component (red), sweet wines exhibit significantly higher values compared to fortified wines, whereas the b* component is lower, indicating a greater presence of the blue component ( Arcari et al., 2013). As a result, sweet wines display a redder hue than fortified wines. Furthermore, considering the two grape drying techniques, wines produced from grapes dried in a cold room (CR) exhibit a more intense colour than those obtained from grapes dried at room temperature (RT).

Table 3. Chemical parameters of wines.

Data are the means (± standard deviation) of 3 wine reps using 3 WineScan ^TM analyses. Different letters in each row indicate statistical differences ( p ≤ 0.05).

The discrepancy between Sweet and Fortified wines was confirmed by the colour difference value (ΔE* _ab) (Table 4). The results indicate that Sweet RT and Sweet CR exhibit markedly different colours compared to Fortified RT and Fortified CR, with an ΔE* _ab value greater than 7.00. Conversely, a perceptible difference in the human eye (ΔE* _ab = 2.84) was observed between Sweet RT and Sweet CR. However, this was not the case for the Fortified RT and Fortified CR, which showed a lower colour difference (2.06), indicating no perceptible difference to the human eye ( Bianchi et al., 2025a; Martínez et al., 2001).

Table 4. Colour difference (ΔE* _ab) among the wines.

ΔE* ab	Sweet RT	Sweet CR	Fortified RT	Fortified CR
Sweet RT		2.84	7.05	7.53
Sweet CR			9.21	9.60
Fortified RT				2.06
Fortified CR

These data were confirmed by HPLC analyses (Table 5) where we observed the almost disappearance of non-acylated anthocyanins in Fortified wines. Moreover, the other class of polyphenols diminished significantly in Fortified wines. Dehydration at cold temperatures maintained a higher concentration of the different classes, both in Sweet and Fortified wines. The reason for these higher values for wines from cooled, dehydrated grapes is due to a higher concentration measured in grapes after dehydration at low temperatures. As for the significantly lower content in Fortified wines, this is due to the lack of extraction from the solids, as a filtration step was applied. In contrast, extraction occurred during fermentation for the other wines.

Table 5. Single phenolic compounds (mg/L) in wines.

Data are the means (± standard deviation) of 3 wine reps. Different letters in each row indicate statistical differences ( p ≤ 0.05). n.d. = not detected.

Regarding VOC analysis (Table 6), a significantly lower amount of VOCs was measured in the Fortified wines across almost all compound classes: an 8–10 fold decrease for acids and a 20–50 fold decrease for alcohols, which are the two most abundant classes. The higher content in VOCs of Sweet wines is likely due to the fermentation process that is lacking in fortified wines. Only aldehydes were higher in Fortified wines, especially benzaldehyde, benzeneacetaldehyde, and syringaldehyde in CR wine. Benzaldehyde and benzeneacetaldehyde have a floral-fruity aroma ( Kuang et al., 2022) and are mainly produced during grape exposure to the sun ( Liu et al., 2024); their highest presence in Fortified wines, could be due to a greater extraction by ethanol from small debris in the must. It is known the affinity of ethanol for lignin ( Hamzah et al., 2021). The presence of syringaldehyde could be the result of lignin degradation due to alcoholic maceration, especially because the torque press could have released a small portion of lignified and dried stem which was not blocked by the filter ( Cernîşev, 2017). Acetic acid was significantly reduced in Fortified wines and this was expected as aforementioned.

Comparing the wines from CR or RT dehydrated grapes, in Fortified wines, except for terpenes where the values were similar, the CR wines had higher concentrations in almost all other classes. In contrast, in Sweet wines, only the classes of alcohols and esters showed significantly higher content in the RT wines.

In Fortified wines, the CR sample had a higher content mainly due to acetic acid, phenylacetic acid, and palmitic acid whereas in Sweet wines the content of acids was similar in the two dehydration conditions. Even classes of alcohols and esters were in higher content in Fortified CR than in Fortified RT, meaning that the lower temperature allowed for preserving VOCs. However, these values were significantly lower than those from Sweet wines where the RT sample had a greater concentration of alcohols and esters. The reason for a higher content of alcohols in Sweet RT is mainly due to phenethyl alcohol and 3-Methyl-1-butanol as a result of amino acid catabolism, which could be due to the higher presence of amino acids in dehydrated grapes at room temperature (optimal temperature for the metabolism) ( Ossola et al., 2017). Esters, and specifically ethyl palmitate, ethyl stearate, ethyl oleate, ethyl linoleate, and ethyl linolenate were higher in RT wines indicating a membrane lipid catabolism, which could suppose a greater fatty acid concentration, confirming what aforementioned for amino acids. Finally, terpenes were significantly higher in Sweet wines and this result could be due to the presence of solid must in fermentation, thus a greater extraction. In conclusion, regarding VOCs, grape dehydration at room temperature results in a more complex aroma due to fermentation, making this technique suitable for sweet wine production. On the other hand, if fortification is the goal, dehydration at low temperatures is preferred.

PLS-DA stands as a robust multivariate analysis method widely employed in chemometrics to unravel relationships between input variables and categorical output variables. The score graph delineates the sample distribution within the space of latent variables (LVs), with each point signifying a sample's disposition relative to the LVs derived from PLS-DA. Here, 2 LVs were chosen to optimise the separation between different wine types, achieving 95 % explained variability (Figure 2).

Figure 2. Score plot (LV1 vs LV2) of the Partial Least Squares Discriminant Analysis (PLS-DA) model built from the data obtained from the measurements with the electronic nose (the numbers 1, 2, 3, indicate the three bottles used for the analyses).

Notably, the score plot vividly illustrates the separation between the Fortified and Sweet wine samples, transcending both CR and RT conditions. The clear segregation of these two groups into opposing quadrants underscores significant aromatic differences between the samples. Sweet wines were also discriminated against for the two dehydration methods while Fortified wines were not, suggesting a high similarity between the two Fortified wines. The variable important in projection (VIP) gives information on the sensors responsible for the segregation of the scores (Figure 3). Notably, sensor 7 emerges as a key discriminator, particularly influential in delineating Sweet RT wine from its counterparts. Conversely, sensors 2, 3, 4, and 6 play a more prominent role in distinguishing Fortified wine. Particularly noteworthy is the association between sensor 4 (QMB4) and certain aromatic compounds, including volatile phenols and furan compounds ( Martínez-García et al., 2021; Saevels et al., 2004), as corroborated by prior studies employing a similar QMB-Enose configuration. Similarly, sensor 7 has been linked to aldehydes and alcohols in existing literature ( Martínez-García et al., 2021; Saevels et al., 2004), suggesting that these classes have particular significance in discriminating between wine types. Aldehydes and alcohols were indeed different in the different wines with Fortified wines having lower content of some alcohols (such as 3-methyl-1-butanol and phenylethanol) and Sweet RT wines containing higher concentrations of 3-methyl-1-butanol and phenethyl alcohol.

The statistical robustness of the PLS-DA model is also evaluated through a confusion matrix, along with related efficiency and robustness indices in calibration and cross-validation (Table S1). During calibration, the model achieves 100 % accuracy in identifying the four classes. In the cross-validation phase, the model maintains good statistical efficiency in discriminating the sweet wines (with a maximum error of 0.08). On the other hand, misclassifications were observed for Fortified wines with CT samples assigned to the Fortified RT class, reinforcing the similarities already observed in the score plot. Even though these two wines presented different analytical aromatic profiles, the slight discrepancies between GC-MS and E-nose measurements are not completely unexpected. Hence, GC-MS gives the precise quantification of specific VOCs present in the analysed matrix but it does not take into account chemical interactions, combinations, and synergy between different molecules ( Tarì et al., 2024). Each sensor in the E-nose array responds differentially to different VOCs, creating a characteristic fingerprint of the sample’s volatile composition. The combined sensor responses generate a dataset representing the overall VOC profile ( Modesti et al., 2024).

QMB-based e-nose has been already used to discriminate wines from different varieties and vintages ( Alfieri et al., 2024; Harris et al., 2023), to identify off-flavours formation ( Cynkar et al., 2007), and to differentiate different sweet wines ( García-Martínez et al., 2011), highlighting its potential. However, other studies suggest that the E-nose systems still present important limitations, especially low selectivity, which make it hard to discriminate different samples when the aromatic fingerprints are not so different ( Celdrán et al., 2022). In this context, it can be hypothesised that the addition of ethanol in fortified wines complicates the E-nose's ability to distinguish between the two wines, as ethanol itself may mask other differences caused by varying dehydration rates. Hence, ethanol generates a strong signal on sensor arrays, which can impair the ability of E-noses to discriminate between different aromas. Moreover, the presence of ethanol can decrease the volatility of other aroma compounds, further complicating the discrimination process ( Ragazzo-Sanchez et al., 2006).

Figure 3. Effect of variables (loadings, QMB sensors of the Nose) on the pattern recognition discrimination of the PLS-DA operated on the acquisitions of the headspaces at the electronic nose. Influences are evaluated as a projection of the VIPs (variable importance projections) generated by each variable (QMBs) and evaluated based on an efficacy threshold (equal to 1).

The PLS-DA obtained with the transformed spectra (absorbance = log 1/T), described 97 % of the total variability across the first two latent variables (LV1 82.15 % and LV2 15.25 %) (Figure 4).

In the score plot, the Sweet wines cluster together regardless of the CR or RT conditions, suggesting a high similarity between them. On the other hand, Fortified wines exhibit clear separation from each other, occupying opposing quadrants. In NIR spectroscopy, spectral absorbances are directly linked to fundamental molecular vibrations (overtones and combination bands) of organic functional groups such as C-H, N-H, O-H, and S-H ( Rouxinol et al., 2022). The vibrations of chemical bonds in complex matrices occur at specific frequencies, determined by the mass of the atoms, the stiffness of the bonds, and vibrational coupling effects. As a result, most biochemical and chemical species in wine exhibit characteristic absorption bands in the NIR spectral regions. Therefore, the analysis of absorbance peaks in the VIP plot (Figure 5) allows for the identification of significant spectral ranges responsible for the discrimination between different wines, enabling some important considerations of the organic molecules involved ( Cozzolino, 2015).

Figure 4. Score plot (LV1 vs LV2) of the Partial Least Squares Discriminant Analysis (PLS-DA) model constructed from data obtained from NIR-AOTF spectral acquisitions (the numbers 1, 2, 3, indicate the three bottles used for the analyses).

For instance, in the Fortified CT wine, the first significant absorption bands have been observed in the spectral region around 1200 nm as well as around 1400 nm, which can be attributed to sugar-related absorption bands ( González-Caballero et al., 2010). Moreover, the higher absorption observed in the spectral portion related to 1400–1450 nm, can be associated with water molecules ( Cozzolino et al., 2006), due to the first overtone of O-H stretch and the O-H asymmetric stretching and bending combination ( Osborne et al., 1993). The higher absorption in the spectral region around 1200–1400 nm is especially relevant, as fortified wines consistently show elevated concentrations of alcohol, from the addition of ethanol, and higher sugars, due to the lack of fermentation. Thus underscoring the capability of NIR spectroscopy to detect the vibrational modes of these molecules. A minor spectral modification can be also observed between the 2000–2200–2400 nm portion, potentially assignable to phenol and condensed tannin ( Cozzolino et al., 2004; Dykes et al., 2014; Zhang et al., 2008). This spectral feature well correlates with the differences observed in anthocyanins and polyphenols observed in the wines (significantly lower in Fortified wines compared to Sweet wines).

Notably, the constructed model demonstrates good performance in class assignments, achieving 100 % accuracy and error equal to 0 in both calibration and cross-validation phases (Table S2). The robustness of the model underscores the reliability of NIR in accurately discriminating the diverse classes. The excellent assignment pattern and the observed absorption bands effectively correspond to compositional behaviours, such as those related to sugars and polyphenols. However, the NIR approach, while effective in class discrimination, can be complex due to the overlapping interactions between absorption bands and the intrinsic characteristics of the wine matrix, making it challenging to establish a direct and precise link to the concentrations of specific compounds ( Littarru et al., 2025).

Figure 5. Effect of variables (loadings, wavelengths) on the pattern recognition discrimination of the PLS-DA operated on spectral acquisitions of NIR-AOTF. Influences are evaluated as a projection of the VIPs (variable importance projections) generated by each variable (wavelengths) and evaluated based on an efficacy threshold (equal to 1).