Learning algorithms for identification of whisky using portable Raman spectroscopy.

Reliable identification of high-value products such as whisky is vital due to rising issues of brand substitution and quality control in the industry. We have developed a novel framework that can perform whisky analysis directly from raw spectral data with no human intervention by integrating machine learning models with a portable Raman device. We demonstrate that machine learning models can achieve over 99% accuracy in brand or product identification across twenty-eight commercial samples. To demonstrate the flexibility of this approach, we utilized the same algorithms to quantify ethanol concentrations, as well as measuring methanol levels in spiked whisky samples. To demonstrate the potential use of these algorithms in a real-world environment we tested our algorithms on spectral measurements performed through the original whisky bottle. Through the bottle measurements are facilitated by a beam geometry hitherto not applied to whisky brand identification in conjunction with machine learning. Removing the need for decanting greatly enhances the practicality and commercial potential of this technique, enabling its use in detecting counterfeit or adulterated spirits and other high-value liquids. The techniques established in this paper aim to function as a rapid and non-destructive initial screening mechanism for detecting falsified and tampered spirits, complementing more comprehensive and stringent analytical methods.