Development of analytical “aroma wheels” for Oolong tea infusions (Shuixian and Rougui) and prediction of dynamic aroma release and colour changes during “Chinese tea ceremony” with machine learning

Highlights

• Type B (Rougui) had mostly higher aroma intensity than type A (Shuixian).
• Innovative use of GC-O-MS generated a specific analytical “Aroma Wheel”.
• APCI-MS/MS showed dynamic changes in aroma release during multiple tea infusions.
• A progressive loss of aroma and colour for seven brews were modelled by MPR and LSTM.
• A minimal release of aroma and colour was found after the fourth brew.

Abstract

The flavour of tea as a worldwide popular beverage has been studied extensively. This study aimed to apply established flavour analysis techniques (GC–MS, GC-O-MS and APCI-MS/MS) in innovative ways to characterise the flavour profile of oolong tea infusions for two types of oolong tea (type A- Shuixian, type B- Rougui). GC–MS identified 48 aroma compounds, with type B having a higher abundance of most compounds. GC-O-MS analysis determined the noticeable aroma difference based on 20 key aroma compounds, facilitating the creation of an analytical “Aroma Wheel” with 8 key odour descriptors. APCI-MS/MS assessed real-time aroma release during successive brews linked with the “Chinese tea ceremony” (Gongfu Cha). Multivariate Polynomial Regression (MPR) and Long Short-Term Memory (LSTM) network approaches were applied to aroma and colour data from seven successive brews. The results revealed a progressive decline in both colour and aroma with seven repeated brews, particularly notable after the fourth brew.

1. Introduction

Tea, made from Camellia sinensis plant, is the second non-alcoholic beverage most consumed in the world after water (Chaturvedula & Prakash, 2011). The world tea production was about five million tonnes in 2013 with almost two million tonnes only for China, the world's first largest producer (Chang, 2015). In general, flavour properties are the most important characteristic of a food product for consumers and are particularly relevant for tea, for which aroma is one of the most important selection criteria (Feng et al., 2019).

Based on the processing methods, six types of tea have been defined: white, yellow, green, dark, oolong, and black teas (Feng et al., 2019). Tea is generally divided into three categories according to the degree of fermentation: green tea is unfermented, oolong tea is semi-fermented with an oxidation level between 10 % and 70 % (Chen et al., 2011) and black tea is fermented. For Oolong tea, the steps of the process are: cultivation, plucking, withering to allow the leaf to lose moisture, leaf disruption to reduce the size of tea leaves, fermentation during which various chemical and biochemical reactions occur, firing, rolling and drying which leads to the end of the fermentation reactions and lowers the water activity content. A wide range of oxidation (10–80 %) in the Oolong tea, particularly controlled enzymatic oxidation, led to distinct sensory characteristics and the effect of processing and chemical composition on oolong tea quality was reviewed by Ng et al. (2017).

The tea culture is particularly important in China, and it has evolved in terms of tea preparation, instruments used and how to serve it. The “Gongfu cha” ceremony is one of the famous Chinese tea ceremonies, where oolong tea is commonly used. Through this ceremony, the same tea leaves are used to brew multiple times, and tea drinkers are encouraged to explore the changes in perceived flavours between the multiple infusions. However, there is a lack of studies to illustrate the changes in aroma release during these successive brews.

The most common analytical technique to analyse tea flavour is chromatography-mass spectrometry GC–MS (Guo et al., 2019). After the identification of volatile compounds by GC–MS, lots of studies also conducted GC–MS Olfactometry (GC-O) analysis when the compounds separated by GC column can be sniffed by human nose to determine the aroma-active compounds (Chen et al., 2019; Sasaki et al., 2017; Song & Liu, 2018). GC-O-MS analysis combines GC-O and GC–MS using the split column, with one end connected to the olfactory detector port and another linked with a mass spectrometry detector. Atomspheric pressure chemical ionisation – mass spectrometry (APCI-MS/MS) technique is one of the online analysis tools to monitor flavour release, which has been applied to different food products like olive oil (Genovese et al., 2018), apple juice (Gan et al., 2014) and confectionaries (Yang et al., 2010), but has never been reported in tea flavour, particularly associated its dynamic changes during tea ceremony.

In addition, sensory wheels or flavour wheels are generated by Quantitative Descriptive Analysis (QDA) using trained panellists, and they normally provide comprehensive sensory profiling (e.g., aroma, taste and mouthfeel) of certain products (Meilgaard et al., 2007). A sensory flavour wheel has been generated for different types of tea, such as herbal tea- rooibos (Koch et al., 2012), honeybush tea (Theron et al., 2014) and barley tea (Goto et al., 2019).

In this study, the first objective was to develop a GC-O-MS method to create an “Aroma Wheel” of oolong tea, specifically for two types of oolong tea (type A: Shuixian, type B; Rougui). The flavour and sensory properties of these two varieties were reported separately in previous studies, such as studies by Wu et al. (2022), and Lin et al. (2024) focused on Shuixian, and Liang et al. (2024) and Wu et al. (2024) studied Rougui. In our study these two varieties, grown in the same region and processed under the same conditions, were compared. The second objective was to assess the real-time aroma release during multiple tea infusions by APCI-MS. The final objective of this study was to employ two machine learning approaches- the Multivariate Polynomial Regression (MPR) model and the Long Short-Term Memory (LSTM) network to model and predict the changes of aroma and colour throughout successive tea infusions. MPR is a type of regression analysis that models the relationship between multiple independent input variables and the dependent output variable as an nth-degree polynomial (Shrivastava et al., 2023). LSTM is a variant of Recurrent Neural Network (RNN) that is capable of learning long-term and predicting future values based on previously observed values (Qiao et al., 2023).

2. Material and methods

2.1. Chemicals

N-Alkanes (C6 to C20) and methanol were purchased from Sigma-Aldrich and used to prepare a solution in hexane to calculate linear retention indices (RI). The saturated salt solution was made with sodium chloride (Fischer Chemical) and pure water from a water purification system (Purite Select Ondeo, Richmond Scientific). 3-Heptanone was purchased from Sigma-Aldrich and diluted at 0.001 % to be used as an internal standard in methanol for GC–MS analysis. Linalool, geraniol, hexanal, 6-methyl-5-hepten-2-one, phenylacetaldehyde, trans-2-octenal, (Z)-3-hexen-1-ol and trans-2-hexenal were purchased from Sigma-Aldrich and used as APCI standard.

2.2. Tea samples

Two types of oolong tea were provided by Xiangjiangchaye Co. Ltd., China: “Shuixian” (translated as “daffodil”) labelled as Type A, and “Rougui” (translated as “cinnamon”) labelled as Type B respectively in this study. They were both growing in the same region and samples were collected from the production years of 2018 and 2019, following the same processing methods for each type (both at 60–70 % fermentation ranges). Three batches of each type per year were provided and were sealed in their original packaging, then stored in a cool and dry place. All the batches from both two years of production were used for GC–MS analysis and sensory Quantitative Descriptive Analysis. The year 2019 batches of tea were used for GC-O-MS analysis and tea ceremony-associated APCI-MS and sensory analysis.

2.3. GC–MS analysis

Tea leaves (5 g) were weighed into a 250 mL beaker and then infused for 5 min with 110 mL of boiled pure water. During the infusion time, beakers were sealed with aluminium foil. The brew was then filtrated with a strainer and introduced into a 200 mL vial. To minimise aroma loss after brewing, vials were immediately placed for 5 min in an ice bath (∼ 40 °C) and then 10 min at room temperature (∼20 °C). A saturated salt solution (2 mL NaCl) was added at room temperature to the tea brew (3 mL) in a crimped amber vial, and the internal standard of 3-heptanone (40 ppb) was added to calculate the relative abundance of the compounds of interest present in the headspace, as described previously (Liu et al., 2019). Four replicates for each type of tea were analysed in a randomised order for GC–MS analysis.

GC–MS analysis was carried out using a GC (Trace 1300, Thermo Scientific, USA) coupled with a Single Quadrupole Mass Spectrometer (Thermo Scientific, USA). The column used is a ZB-WAX (30 m × 0.25 mm × 1.00 μm, Phenomenex Inc., Macclesfield, UK). SPME procedure was carried out for GC–MS and GC-O analyses. A 50/30 μm DVB/CAR/PDMS SPME Fibre (Supelco, Sigma Aldrich, UK) was exposed to the sample headspace for 30 min at 70 °C with agitation and then desorbed for 1 min.

Helium was used as the gas carrier at a flow rate of 1.00 mL/min and a constant pressure of 30 psi (2.07 × 10⁵ Pa). The oven temperature was initially at 40 °C, held for 2 min, then increased to 240 °C at a rate of 6 °C/min, maintained for 5 min. Electron ionisation was selected as the ionisation mode with a mass range scan of 35–300 amu (atomic mass unit) and a scan time of 0.2 s. This GC–MS method was built on the method reported by Yang et al. (2016). All components are semi-quantified against the internal standard of 3-heptanone (40 ppb) that was added to calculate the abundance of the compounds.

The limit of detection (LOD) and the limit of quantification (LOQ) of every compound was based on the Signal to Noise Ratio (S/N), and the LOD was set at S/N > 3 and LOQ was at S/N > 5. Three identification methods were used, including i) reference libraries (NIST/EPA/NIH Mass Spectral Library, National Institute of Standards and Technology, Gaithersburg, MD), ii) the linear retention index (Ullrich and Grosch, 1987) from literature (WILEY07 database) and iii) the LRI of the standard solution. The LRI values were calculated by running the homogenous series of n-alkanes (from C6 to C26) under the same GC conditions (Sun et al., 2020).

2.4. GC-O-MS analysis

The GC-O-MS analysis was realised with the same GC–MS used in GC–MS analysis with an additional sniffing port. The column used is a ZB-WAX (30 m × 0.25 mm × 1.00 μm, Phenomenex Inc) and was split between the MS and the sniffing port. The same GC–MS operation conditions were applied in the GC-O-MS analysis as in the previous GC–MS analysis.

Following a similar method published by Kirkwood et al. (2024), the Olfactometry panel was recruited at the University of Nottingham (Sutton Bonington, UK) and it consisted of ten panellists (aged 21–35). The training was provided in the first session, where each panellist practised how to record retention time, intensity value and odour description when they detected a note. The intensity was measured by a 3-point scale: 1 for weak, 2 for medium and 3 for strong. To help the panel with the aroma description, a list of the characteristic scent of tea was given before the session. Each panellist sniffed both types of tea twice at different sessions over 4 days.

2.5. APCI-MS/MS analysis

An MS-Nose interface (Micromass, Manchester, UK) fitted to a Quattro Ultima mass spectrometer (Waters Corporation, Milford, MA) was used to analyse the static headspace of tea samples. The static HS analytical method for APCI-MS analysis was reported previously (Yang et al., 2020). The tubing used was a 50 m length and 0.32 mm internal diameter Zebron deactivated tubing (Phenomenex, USA). The source temperature was 50 °C, the transfer line temperature was 120 °C, the dwell time was 0.1 s and the flow rate was around 30 mL/min. MassLynx software (MassLynx v3.2, Micromass Ltd., Manchester, UK) was used to integrate peaks, recover their maximum intensity and calculate the signal-to-noise ratio.

Based on previous GC-O-MS results, seven aroma compounds were selected to be the target for APCI-MS/MS analysis (Appendix Table A.1). To develop the analytical method, tea infusion samples were analysed firstly in “Full Scan” mode to monitor ions of mass to charge ratio 20–200. A standard solution of the individual compound was analysed in “Selected Ion Recording (SIR)” mode, followed by “Multiple Reaction Monitoring (MRM)” mode. Optimised cone voltage and collision energy were applied to achieve the highest signal for each compound.

During the tea ceremony procedure, boiling water is normally used to preheat the flask and to rinse the tea leaves for 5 s, so a preliminary analysis was conducted to evaluate the impact of a flask pre-heating before the first brew. A temperature probe was placed in the tea infusion to monitor its temperature throughout the whole process (Fig. A1). The temperature of the first brew infusion was reduced from ∼65 °C to ∼45 °C during 5 min infusion period without preheating the flask, and other infusions from the rest of the six brews showed similar changing patterns from ∼65 °C to ∼50 °C during the infusion period. Alternatively, pre-heating with 100 mL boiling water was found to maintain similar temperature changes in the first brew to the successive brews. Since temperature could have a significant impact on aroma release from the infusion, pre-heating the flask was proved to be an effective way to minimise this impact between the first brew and the rest of the brews in this study.

For each acquisition, tea leaves (1 g) were weighed into a 200 mL Schott flask, and the schematic diagram of this experiment was illustrated in Appendix Fig. A2. After pre-heating the flask with 100 mL boiling water for 5 s, another 100 mL boiling water was added to the tea leaves in the flask for 5 min infusion, and its aroma released into the headspace was analysed by APCI-MS/MS for 20 s at every minute during the infusion period. Then this tea infusion was decanted from the flask, which was used for colourimetric analysis. This infusion process was repeated seven times using the same tea leaves, but fresh boiling water (100 mL) was added every time and then decanted after 5 min infusion. Both types of tea had seven successive brews with six replicates to collect data on the ion intensity. The maximum ion intensity (Imax) of each target ion at every brew and the time to reach the maximum intensity (Tmax) were recorded and analysed. Six replicates of each type of tea were used for both APCI-MS and colourimetric analysis.

2.6. Colourimetric analysis

The colour of multiple tea infusions collected from APCI-MS/MS was analysed by a spectrophoto-colourimeter (Lovibond LC 100, model RM 200, The Tintometer Ltd., UK) using the following parameters: D65 illuminant and 10° observer angle. CIELAB parameters (L*, a* and b*) were measured, where L* is the lightness value (0 for dark and 100 for light), a* is the red/green axis and b* is the yellow/blue axis. Another three parameters (the chroma value C, the hue angle h and the colour difference ΔE) were calculated from L, a* and b* based on the following equations (HunterLab, 2008):

\(C = \sqrt{a^{\ast 2} + b^{\ast 2}}h = \arctan \left(\frac{b^{\ast}}{a^{\ast}}\right)\)

\(\mathrm{ΔE} = \sqrt{ΔL^{\ast 2} + Δa^{\ast 2} + Δb^{\ast 2}}\)

2.7. Sensory analysis

The first sensory test was based on the Quantitative Descriptive Analysis (QDA) with twelve trained panellists recruited. The preparation of tea infusion followed similar methods as the samples used for GC-O-MS analysis. Eight attributes (green, flora, sweet/fruity, nutty, roasted, potato, mushroom and woody) were selected. The panellists were instructed to sniff the tea infusion without drinking it, using an established method (Yang et al., 2011). Panellists rated the intensity of each attribute by marking a linear scale with anchors marked 0 and 10, where 0 = none and 10 = the highest possible intensity. Assessment of both types of samples was undertaken in three replicated sessions over three days, and the panellists' scores from these sessions for each attribute were averaged.

The second sensory evaluation was conducted by seven trained panellists using both type A and B teas produced in the year 2018 and year 2019. Tea infusion with seven successive brews was prepared in a similar method as in APCI-MS analysis: 1 g tea leaves (type A or B) with 100 mL boiling water, and 10 mL of the tea infusion was poured into the glass container after 5 min of each brew. Every panellist was instructed to sniff the headspace of each brew and score their perceived intensity of the nutty/roasted and flora/fruity on a scale of 0 to 10, a similar scale as the first sensory test.

2.8. Statistical analysis

All the statistical analysis has been done using SPSS (IBM® SPSS® Statistics version 25) and XLSTAT Software ©-Pro (2019.3.1, Addinsoft, Inc). Multivariate analysis of variance (MANOVA) and Tukey's post hoc test was conducted by SPSS. The level of significance was set at p < 0.05. Principal component analysis (PCA) was conducted by XLSTAT to compare the volatile differences between the two tea samples.

2.9. Machine learning approaches

Multivariate Polynomial Regression (MPR) was applied to model the non-linearity relationship between input brew sequence and output ion releases. The model f(x) for each ion has the following form:

\(f\left(x\right) = w_1\frac{1}{x} + w_2\frac{1}{x^2} + w_3\frac{1}{x^3} + b\)

where w₁, w₂, and w₃ are parameters of each polynomial option and is the bias. They are learned by using the collected ion numbers in the brew sequences. To avoid model overfitting (i.e., too complex model that impedes generality), the polynomial option that had an insignificant gradient by setting the corresponding w to zero was removed. The final function is the one that minimises the root mean square error (RMSE) of the real ion data points with the restriction in place. RMSE is the square root of the mean of the squared error for all predictions:

\(\mathrm{RMSE} = \sqrt{\frac{1}{n}{\sum}_{i = 1}^n\left(y_i- \hat{y}_i\right)^2}\) , where is the true value, is the predicted value, and n is the number of observations available in the dataset (Christie & Neill, 2022).

In time-series prediction, LSTM networks are ideal for this task because they can remember information for long periods, and are thus able to capture patterns that span several time steps (Abbasimehr & Paki, 2022). Trained on real-time tea aroma release obtained through APCI-MS/MS analysis, LSTM is employed to predict the number of ions released in the next brew using data from previous brews. Before learning parameters for LSTM, a data selection approach was designed to select the best-performed data points. Specifically, the data points were ranked by their autocorrelation to their immediate preceding data point and only used the top-ranked data for training. The LSTM structure used has 64 stacks of LSTM hidden units and one fully connected for the predicted output. It was trained with each tea type's ion's data points resulting in 20 different models for two tea types, 5 ions and different brew series (1st to 6th brew and 1st to 5th brew). Similar machine-learning approaches were applied to colour data collected from each brew.

Once the models had been trained, they were used to predict the next brew's specific Ion release amount of the tea based on the previous ion data points. For the model trained on the 1st to 6th brew, the 7th brew was predicted. For the model trained on the 1st to 5th brew, the 6th and 7th brew will be predicted. For the latter model, the 7th brew will be calculated based on the predicted values of the 6th brew.

3. Results and discussion

3.1. Identify aroma differences using GC–MS analysis

A total of 46 aroma compounds were identified by GC–MS for both types of tea and classified into 8 functional groups (Table 1), including 14 aldehydes, 12 alcohols, 11 ketones, 3 esters, and 5 heterocyclic organic compounds. Most of the compounds identified have been already reported in oolong tea studies (Hu et al., 2018; Ma et al., 2018; Zeng et al., 2017) or other types of tea (Chen et al., 2019; Wang et al., 2018).

Food Chemistry, Volume 464, Part 1, 2025, 141537: Table 1 (truncated). Aroma compounds identified by GC–MS analysis.

To compare overall aroma profiles across these two types of tea, Principal Components Analysis (PCA) was applied for all 46 aroma compounds. Two production years for the same tea types were compared, as shown in Fig. 1 i) and ii). Despite the year of production, the first principal component (PC1) effectively illustrated the volatile differences among samples, that is, type B had a higher level for most aroma compounds than for type A. Statistically, 34 out of 46 aroma compounds detected in 2019 had significantly higher relative headspace for type B than for type A (p < 0.05). Only 3 compounds (2-methylpentanal, indole and tea pyrrole) in year 2019 were found at significantly higher levels in type A than type B (p < 0.05).

Food Chemistry, Volume 464, Part 1, 2025, 141537: Fig. 1. Principal Components Analysis (PCA) of 48 aroma compounds identified by GC–MS analysis for 3 batches of tea infusion for type A and type B harvest in (i) year 2018 and (ii) year 2019.

GC–MS coupled with SPME was shown to be a useful and effective technique to get a comprehensive list of aroma compounds in the matrix and to demonstrate the differences between samples. However, aroma compounds that showed higher peaks at higher levels in the products did not necessarily indicate their importance and contribution to the overall aroma profile. Some aroma compounds might present at a trace level but have a very low threshold of detection by the human nose that can lead to a larger contribution to the overall flavour. Therefore, linking with human olfactory perception through odour port, GC-O analysis that had been used to identify key aroma compounds in previous tea studies (Sasaki et al., 2017) was also applied in this study.

3.2. Create an analytical “aroma wheel” using GC-O-MS analysis

Twenty compounds were identified through GC-O-MS analysis by 10 panellists (Table A.2). This is an adapted method from OSME (McDaniel et al., 1989), which allowed a direct estimation of the odour intensity from the sample extract using a reasonable number of panellists, instead of using multiple dilutions in other GC-O-MS methods like AEDA (Ulrich & Grosh, 1987) or CHARM (Acree et al., 1984). However, the 15 cm long category scale in OSME could be challenging for the panellists to score within a short response period (e.g., 2–5 s) when they sniffed an odour, so the scale of 1–3 to indicate weak, medium and strong odour was used in this study. The selection criteria were improved as if the note of an aroma compound was recorded by at least five panellists, with adjacent retention time and a similar descriptor, this compound was included as a key compound. These compounds were classified into 8 main odour categories based on their odour descriptors collected from the panellists, which were: green, woody, mushroom, nutty, roasted, potato, floral, and fruity/sweet. Similar descriptors were also found in other studies for oolong tea (Ma et al., 2018; Zeng et al., 2017; Chen et al., 2020; He et al., 2023).

In Table A.2, perceived frequency (Freq) recorded the number of panellists who had identified this compound (scale 5–10), average intensity (AvI) was defined as the mean of all the intensity from all the panellists (scale 1–3), and the total intensity value (TIV) was calculated as the frequency (Freq) times average intensity (AvI). High TIV value for a compound (e.g., (Z)-hex-3-en-1-ol with TIV ≥ 20) indicated that it can be perceived by a large portion of panellists (e.g., Freq =10) at a relatively high intensity (e.g., AvI = 2 and 2.4 for type A and B respectively). Some compounds (like 2-phenylacetaldehyde) were not perceived by all the panellists (Freq =9) but had high odour impact (TIV >20) for both types of tea, because the AvI is high (2.4) indicating the perceived intensity was very strong.

Additionally, four compounds (1-ethyl-pyrrole, nonanal, isophorone, (3E,5E)-octa-3,5-dien-2-one) showed in Table A.2 were not perceived as key compounds in type A, but they were identified in type B by more than 7 panellists, which could contribute to woody, floral and green note in Type B. Apart from most compounds that had higher TIV in type B than type A, 2-methyl butanal and hotrienol that linked with green, woody and floral notes had higher total intensity in type A.

To visualise the aroma differences between the two types of tea, an analytical “Aroma Wheel” was created based on the TIV results of all 20 key aroma compounds from GC-O analysis and their classified odour description (Fig. 2 i)). This “Aroma Wheel” illustrated that two types of tea had distinct profiles, and type B with a generally larger area than type A had stronger green, woody, mushroom and roasted notes. Whilst, type A might have a similar floral, sweet and fruity note to type B, but it could be slightly floral due to its slightly higher value for hotrienol. The main difference could be linked to the relative abundance indicated by the PCA plot: type B had a higher concentration in most aroma compounds than type A's, so key aroma compounds in type B were perceived with stronger intensity than type A.

Food Chemistry, Volume 464, Part 1, 2025, 141537: Fig. 2. Flavour profile of tea infusion comparison between type A (blue solid line) and type B (orange dashed line) based on (i) “Aroma Wheel” of 20 key aroma compounds identified by GC-O-MS analysis; (ii) Simplified “Aroma Wheel” of 8 key odour descriptors from GC-O-MS analysis and (iii) Sensory profile of Quantitative Descriptive Analysis (normalised to Type B data as 100 %). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Additionally, the proportion of each descriptor relative to the global aroma profile could be taken into account, so the relative TIV% contribution of each compound to the sum of TIV of all compounds (as 100 %) was calculated. Then the compounds with similar odour descriptions were combined (e.g., eight compounds were added to give a “Green” note TIV%). Similarly, the relative TIV% values were calculated for eight main descriptors (Fig. 2 ii)). This simplified spider diagram showed that type B had more woody and green notes, but type A showed floral and fruity notes.

The sensory evaluation results (Fig. 2 iii)) showed less apparent differences between these two types of tea when panellists sniffed the headspace above their infusions. Nonetheless, type B was still perceived as more woody and green than type A, as demonstrated by analytical aroma wheels.

Therefore, our study demonstrated the feasibility of creating an “Aroma Wheel” based on GC-O-MS data, and the results highlighted the aroma differences between the two types of oolong tea. The advantage of using GC-O-MS was to focus on the key aroma-active compounds and the resulting “Aroma Wheel” visualised the major flavour differences between products. The novel use of GC-O-MS to create an “Aroma wheel” indicated that a similar approach can also be applied to other food systems in future studies. However, panellists who smelt individual compounds using GC-O-MS might perceive differently when these compounds were combined as a mixture above the headspace of the tea samples. Therefore, this approach can be used as a quick mapping of the flavour profile between products, but it will be useful to have an additional sensory analysis to compare with the analytical results in future studies.

3.3. Monitor real-time aroma release during successive tea brews using APCI-MS/MS

Atmospheric Pressure Chemical Ionisation (APCI) is a real-time headspace analysis technique, and it was the first time used to monitor aroma release that mimics the Chinese tea ceremony in this study (Fig. 3). Using the GC-O-MS results, seven volatile compounds with different physicochemical properties and relatively high TIV values were selected as the target compounds for APCI-MS/MS analysis: (Z)-hex-3-en-1-ol, (E)-hex-2-enal, (E)-oct-2-enal, 6-methyl hept-5-en-2-one, 2-phenylacetaldehyde, linalool and geraniol (Table S1).

Food Chemistry, Volume 464, Part 1, 2025, 141537: Fig. 3. Relative release signals from static headspace above tea infusion during seven successive brews from two types of tea for (i) type A (solid lines) and (ii) type B (dashed lines) measured by APCI-MS/MS with 5 targeted ions (m/z 83, 99, 109, 121, and 137). The signals were normalised to the maximum release signal for each ion as 100 %. Average Imax (iii) and Tmax (iv) for type A and type B during seven successive brews. Error bars are standard errors with different letters (a, b, c, and d) indicating a significant difference at p < 0.05 by Turkey's test.

The limitation of using simple APCI-MS was to distinguish compounds with the same molecular weight, so the MS/MS function was applied to investigate if different precursors and product ions could be used to discriminate them. However, the standard solutions of trans-2-octenal and 6-methyl-5-hepten-2-one showed the same precursor ion (m/z 109) and product ion (m/z 67), so they were classified as one Analytical Group. Likewise, linalool and geraniol with the same precursor ion (m/z 137) and product ion (m/z 81) were also classified into another Analytical Group. The other three compounds (cis-3-hexen-ol, trans-2-hexenal, 2-phenylacetaldehyde) had distinct precursor ions (m/z 83, 99, 121), and they were determined as three individual analytical groups. All five selected analytical groups showed good signals from the standard and tea infusions. Additionally, each ion might not only represent one particular compound as some fragments of other compounds in the tea infusions could have similar m/z values, so instead of using the possible representative aroma compounds, the target ions (m/z 83, 99, 109, 121 and 137) were reported in this study.

The aroma release signal for all six replicates was averaged for each target ion and then normalised to its maximum release as 100 %. The plot of the normalised aroma release intensity against APCI-MS/MS analysis time was shown for type A and type B in Fig. 3 i) and ii) respectively. The first brew started at 10 min and its headspace was analysed for 20 s at every 1 min, then finished at 15 min 20 s. This procedure was also applied to all other brews and replicates. In general, type A and type B showed similar patterns during successive brews, that is, there was a clear reduction in all five targeted ions during seven successive brews. The first brew had the highest signals for all the compounds and then decreased progressively from 2nd brews to 7th brews. Among all five target ions, Ion 137 had the least progressive loss during repeated brews, which remained ∼40 % at the 4th brew, but the other four ions (m/z 87, 99, 109 and 137) had only 10–20 % at the 4th brew. Ion 137 is likely associated with linalool and geraniol (Table A.1), and these two volatiles were the most hydrophobic (the highest Log P) and the least volatile (highest vapour pressure) than other selected compounds. Because of these physicochemical properties, they can be more difficult to extract by water, so multiple extractions are needed to extract. In addition, these hydrophobic compounds are also more easily pushed from the water into headspace once they are extracted, which might lead to a strong signal of Ion 137 in HS and this intensity remained higher in the HS after multiple extractions.

To compare the aroma release difference between type A and B, the average Imax and Tmax of all ions at each brew were calculated (Fig. 3 iii), iv)). In general, type B showed higher Imax and Tmax than type A at each brew based on the results from the selected five ions. Despite the types of tea, a significantly higher Imax was observed in the first three brews (p < 0.05) and no significant difference after 4th brew till the final 7th brew (p > 0.05). It could take longer for ions in type B to reach their Imax than type A at each brew, but this difference gradually diminished toward the final brew.

To the authors' knowledge, this is the first demonstration of the successful application of APCI-MS to monitor the dynamic changes in aroma release during the Chinese tea ceremony. As different varieties of tea can be used and various infusion times could be applied, more studies could be done to compare the optimal diffusion conditions accordingly.

3.4. Evaluate aroma perception during successive tea brews

Additional sensory analysis was conducted using the same brewing methods as APCI-MS/MS analysis. The average sensory intensity scores perceived for floral/fruity and nutty/roasted were recorded from each bew for both types of tea produced in the year 2018 (Fig. 4 i), ii)) and 2019 (Fig. 4 iii), iv)). In general, the aroma perception of both types of tea decreased during the successive brewing process. Despite the year of production, type A mostly scored higher than type B in terms of their perceived aroma intensity, which was different to the Imax results from APCI-MS analysis (Fig. 3 iii)). Interestingly, the difference between these two types of tea was more obvious in the first four brews from APCI-MS results, but perception changes were less apparent during the first two brews and the final six and seven brews. Therefore, APCI-MS data showed that aroma release was reduced significantly during repeated brews, but sensory results indicated that otho-nasal perception differences might be less noticeable between different types of tea. It will be useful to have retro-nasal data in future to compare aroma release during the consumption of tea with the perceived intensity from drinking different successive brews.

Food Chemistry, Volume 464, Part 1, 2025, 141537: Fig. 4. Sensory perception scores of floral/fruity and nutty/roasty for (blue solid line) and type B (orange dashed line) produced in 2018 (i) (ii) and 2019 (iii) (iv) using 3 batches replicates per year and assessed by seven trained panellists. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

3.5. Investigate colour changes during successive tea brews using colourimetric analysis

The colourimetric results of the successive brews for both types of tea infusion were recorded analysed (Table 2). The L* value increased significantly with the number of brews for both types of tea (p < 0.05), which indicated that the lightness of the brew increased. Values of a* and b* decreased with the number of brews significantly (p < 0.05), representing the reduction in red and yellow for both types of tea. The chroma C decreased with the number of brews significantly (p < 0.05), which illustrated that the first two infusions had significantly more saturated colour than Brew 3–5, and the last two brews had less saturated colour (Tukey's analysis, p < 0.05). The hue angle, the association with the colour shade, had no significant difference with the number of brews for type B (p > 0.05), but type A had a significantly different shade between the last brew and the first three brews (p < 0.05). The difference of colour ΔE (ΔE ≥ 2.3 ± 1.3) corresponds to a Just Noticeable Difference by human eyes (Mahy et al., 1994), and the results in Table 2 indicated that there was a significantly perceptible difference for the first four brews as the values were higher than 3.

Food Chemistry, Volume 464, Part 1, 2025, 141537: Table 2. Colorimetric results of both type of oolong tea (type A and B) for each brew (1–7).

In general, with the number of times the leaves were reused, the colour of the infusion became less intense, and the colour differences during the first 4 brews were more perceivable. The molecules mainly responsible for the colour in tea are chlorophylls for the green, theaflavins, flavonols glycosides and carotenes for the yellow, and thearubigins for the red (Chaturvedula & Prakash, 2011). Those molecules were released in the water during infusion, and fewer molecules remained after the leaves were reused multiple times. Both types of tea showed a similar pattern on the reduced colour intensity.

3.6. Modelling of aroma and colour changes during successive tea brews

Both MPR and LSTM models were applied based on the normalised data, which were scaled according to the maximum ion release per ion type in each tea type. This allowed us to compare the ion release patterns on the same scale for different ion types in the same tea type and for different tea types. The final function was calculated using polynomial regression of normalised each ion released for Type A and B with each constant rounded to the nearest 2 decimal points as shown in Fig. 5 (i).

Food Chemistry, Volume 464, Part 1, 2025, 141537: Fig. 5. Linear regression model comparison between type A (blue solid line) and type B (orange dashed line) for (i) 5 targeted ions (m/z 83, 99, 109, 121, and 137) and (ii) colour data (L, a, b). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Type A and type B showed different dynamic release curve: type A showed generally higher release than type B across the seven brews. Similarly, the polynomial regression of normalised colour data was illustrated in Fig. 5 (ii). Compared to type B, type A showed a lower value in colour L changes and remained lower for colour b* changes, whilst not much difference for colour a* for the first six brews.

Additionally, the predictive LSTM models were applied for type A results including all five ions and colour parameters, specifically Fig. 6 (i) and (ii) predicted the 7th brew based on the first six brews and Fig. 6 (iii) and (iv) for the 6th and 7th brews prediction using data from the first five brews. The overall root mean square error (RMSE) of approximately between 0.008 and 0.033 for the 7th brew, and approximately between 0.014 and 0.053 for the 6th and 7th brews. Notably, the LSTM model exhibited superior performance in predicting the 7th brew exclusively. This is most likely due to the difference in quantity on the limited training dataset between the different models, as the more training set a model has, the better the result would be. While encountering challenges in accurately forecasting steep downward and upward movements, the model consistently succeeded in capturing the general trend of ion behaviour. The combination of aroma and colour models from seven successive brews linked with the tea ceremony illustrated that the first 3 to 4 brews had significantly higher aroma and colour intensity. Certain ions (137) showed higher persistency than other ions during successive brews.

Food Chemistry, Volume 464, Part 1, 2025, 141537: Fig. 6. LSTM models from type A tea infusion based on the release of 5 targeted ions (m/z 83, 99, 109, 121, and 137) and colour data (L, a, and b) to predict 7th brew's aroma (i) and colour (ii) and to predict 6th and 7th brew's aroma (iii) and colour (iv).

4. Conclusion

In this study, three flavour analytical techniques (GC–MS, GC-O-MS, APCI-MS/MS) were successfully applied in combination, for the first time, to characterise the aroma profile differences between two types of oolong tea. The innovative use of GC-O-MS to create an analytical “Aroma Wheel” could also be applied in future studies to visualise the overall aroma profile differences between other food products based on their key odour descriptors.

The successful application of APCI-MS/MS, demonstrates for the first time real-time tea aroma release during seven consecutive brews associated with the “Chinese tea ceremony.” Due to different tea varieties and various infusion conditions that could be used in the “Chinese tea ceremony”, this approach could be further applied to other tea varieties to validate the optimal infusion conditions required respectively.

The utilisation of machine learning approaches, specifically the MPR models and the LSTM networks, successfully modelled and predicted the pattern of the aroma and colour changes throughout successive tea infusions. Future studies employing machine learning approaches could explore and predict the dynamic changes of flavour compounds under different processing and storage conditions. These comprehensive and novel approaches may contribute significantly to understanding flavour dynamics and optimising flavour retention in a wide range of food products.