WINETECH Technical Yearbook 2019

method (that will be calibrated), and the appropriate statistical tools to link (calibrate) the two. • In the case of YAN, reference methods range from titration (formol, total YAN), to NOPA (for FAN), enzymatic (for ammonia), and even HPLC (for individual amino acids). From a routine perspective, there are many downfalls when it comes to these methods. The sample preparation is often difficult and time- consuming and may require the use of highly trained personnel, the run time on the required machinery may be quite long, they may make use of expensive and hazardous reagents and they may not maintain the integrity of the sample. • Therefore, the new method should address these shortcomings. As such, the method to be calibrated should have no/minimal sample preparation and require less highly trained personnel, has a short run-time (high throughput), is cheap to run, makes use of no reagents, and possibly maintains the integrity of the sample. For all these reasons, the alternative method proposed for YAN entails the use of IR spectroscopy. • The link between the two methods is done with the help of chemometrics (aka chemical statistics). Chemometrics offer powerful tools that can deal with large amounts of data, can extract useful

Fisher Scientific, Waltham, MA) and the Megazyme™ K-PANOPA (Ireland) for FAN and Enzytec™ Fluid Ammonia (R-Biopharm, Germany) kits. For the IR spectroscopy, three benchtop instruments were tested: MPA FT-NIR (Bruker Optics, Germany), Alpha-P ATR FT-MIR (Bruker Optics, Germany), and WineScan™ FT120 (FOSS Electric, Denmark). Two additional instruments were tested for the cultivar effect task, MicroNIR (VIAVI Solutions Inc., USA) and FieldSpec 4 Standard-Res Spectroradiometer (ASD Inc., Malvern Panalytical, USA). The statistical modelling was done on OPUS v. 7.2 for Microsoft (Bruker Optics, Germany). TASKS To illustrate the feasibility of using IR for YAN determinations, we have considered two practical scenarios. It is known that YAN values can be affected by vintage effect. Therefore, in the first scenario, a model based on previous years’ samples (2016 and 2017) was built and used to predict values for a new vintage (2018). This can correspond to a real case scenario, in which the samples arriving for testing have to be measured using calibrations generated in previous years. The results showed that both the WineScan and the MPA performed the tasks to the required level for quantification. The Alpha-P can be used in this scenario only for screening the new samples, not for accurate quantification.

2) TASKS

3) EVALUATION

1) DATA

Calibration set: Major cultivars (Sauvignon blanc, Chenin blanc, Chardonnay, Shiraz, Merlot, Cabernet Sauvignon) Validation/Test set: Minor cultivars (Marsanne, Rousanne, Pinot gris, Verdelho, etc.)

R 2

, R 2

, RMSEC,

CULTIVAR EFFECT VINTAGE EFFECT

CAL

VAL

RMSEP, RPD CAL , RPD VAL - pre-processing

Spectra + reference values

techniques were used for model optimisation

Calibration set: 2016 + 2017 samples

Validation/Test set: 2018 samples

FIGURE 1. Strategy for the evaluation of the models based on cultivar and vintage effect.

information and can reduce the noise and irrelevant information for the task at hand. The performance of a calibration model is evaluated by certain statistical parameters that indicate in essence, how good a predicted value is, how close to the real one (accuracy), and how well a model can perform when certain tasks are given (robustness).

STRATEGY Settled juice samples (911) were collected over three vintages (2016-2018). A total of 28 cultivars (12 white and 16 red) and 14 grape-growing districts (Sawis) were represented in the set. The reference va l ues for FAN and ammon i a we re generated using the Arena 20XT (Thermo