Adam Gilmore,* Jana Hildreth, and Lyufei Chen
*Horiba Instruments Inc, 9 E View Ct, Flemington, NJ, 08822
(adam.gilmore@horiba.com)

Commercial additive concentrates (CAC) of Teinturier grapes with intensely colored skins and flesh, enriched in malvidin-3,5-diglucoside, are commonly used covertly to enhance color in finished wines from Vitis vinifera sp. grapes. The target CAC concentration is typically at least 0.2% by volume. This study investigates the detection of CAC with A-TEEM spectroscopy using a recently introduced, single-variable adjustment classical least squares (CLS) method known as Gray-CLS (Eigenvector Inc., Solo v9.3.1). The first-principles-based Gray CLS-results are compared to other multivariate regression methods, including partial least squares (PLSR), locally weighted regression (LWR), and extreme gradient boosting (XGBR).  The experiment included several wines from different grape varieties spiked at varying concentrations with a commercially available CAC. A-TEEM measurements were collected under Beer-Lambert linear absorbance conditions at a constant temperature (20°C) using a standard solvent (50% EtOH, pH 2) and 0.45 micron filtration. The model data included both a calibration (cross validation set) of ~80% of the samples and an independent validation set comprising ~20%. The Gray-CLS model yielded a standard deviation (SD) of ~10% of the target CAC (0.2%) for the validation set by optimizing only the general least squares weighting variable of the CLS residuals. PLSR, XGBR, and LWR achieved similar or slightly lower SD values.  However, these methods all required complex and potentially ambiguous optimization of multiple preprocessing variables of the spectral and concentration data blocks and other algorithm-tuning parameters, including the number of latent variables and/or principal components, among others. These methods, prone to under- or over-fitting, are thus potentially unreliable. We conclude the A-TEEM method can be an effective tool to quantify CAC using Gray-CLS to avoid issues with under- and over-fitting multivariate regression models, yielding results relevant to commercial wine quality evaluation.

Funding Support: Horiba Instruments Inc. Internal Funds