There are several conventional methods used to maintain the quality of sparkling wine. For instance, physical and chemical analysis allows specialists to assess the acidity, volume of alcohol and sugar, and other parameters; meanwhile, sommeliers check the product via tasting. And even though these methods provide important insights, they aren’t always reliable: physical and chemical analysis requires expensive equipment, reliable samples, and specific lab conditions, while the results of tasting depend on each sommelier’s subjective opinion.
A more accurate assessment of any liquid can be done with the help of ultrasonic cavitation. Ultrasound causes cavitation bubbles to appear, expand, and contract within the beverage. Unlike “natural” bubbles, which affect the look and taste of the product, these artificial bubbles represent all of its characteristics. The properties of the bubbles depend on the liquid – its alcohol volume, viscosity, and surface tension. The behavior of these cavitation bubbles is recorded on camera and analyzed with the help of computer vision and machine learning technologies. In this manner, researchers at ITMO have already succeeded in classifying water-ethanol solutions by ethanol volume and assessing the quality of petroleum products with varying octane numbers.
However, the combination of ultrasonic cavitation, computer vision, and machine learning has never previously been used to classify the types of sparkling wines – or wine glasses. Thus, researchers from ITMO’s Infochemistry Scientific Center and the University of Wisconsin-Milwaukee (USA) became the first in the world to suggest a new approach and expand the application of these methods.
“Before this, AI wasn’t used in the assessment of sparkling wines for three reasons. Similar alcohol volume in wines can reduce the sensitivity of chemical analysis; the presence of naturally occurring bubbles causes confusion; and the geometry and material of the glass affects the behavior of the bubbles and introduces an additional variable. We’ve managed to get around these limitations. Ultrasonic cavitation creates controllable bubbles regardless of natural aeration; standardized recording and analysis of bubbles allows us to nullify the impact of the glass; and the use of machine learning helps us identify minor differences in the shape and distribution of bubbles – something that can’t be spotted during regular analysis,” says Ilya Korolev, one of the authors of the study and a third-year student at the Infochemistry Scientific Center.
(A) Typical bubbles in a control series (w/out ultrasound) and in a cavitation series (B). Image courtesy of Ilya Korolev
Read also:
ITMO Scientists Train Neural Networks to Determine Solution Concentration Via Air Bubbles
The scientists settled on rose and white sparkling wines, as the components and production technologies of these commonplace drinks differ from one another. For the control study, samples of each wine were poured into vessels made from glass and plastic. Then, the naturally occurring bubbles were recorded on camera before the samples were treated with ultrasound. Changes in the behavior of cavitation bubbles were, too, recorded. Using computer vision and machine learning, the researchers processed 19,953 images and categorized them using AI.
“We were able to prove that bubbles in different types of wine and different containers have visually distinct characteristics that are hard to spot with the naked eye, but can be identified with the help of our algorithms. We segmented the video using the YOLOv8 network, converted the bubbles’ visual characteristics into numbers using the CLIP model and its upgraded version SigLIP, and then used the TabNet classifier for the final sorting. We noticed that the accuracy of classification is vastly improved when the contours of the bubbles are identified using segmentation. This significantly boosts the quality of identified properties,” comments Timur Aliev, the first author of the study and an assistant at the Infochemistry Scientific Center.
(A) Schematic of the experiment. (B) Two types of wine (C) are stored in glass and plastic vessels, then (D) treated with ultrasound. Image courtesy of Ilya Korolev
The new system identifies the type of sparkling wine (rose or white) with 84% accuracy and the type of wine glass (plastic or glass) with 82% accuracy. This invention makes it possible to develop automated, quick, and inexpensive methods for the assessment of quality and authenticity of sparkling wines. Thus, it should prove helpful to production engineers at wine-making facilities, lab researchers, and even consumers wanting to check the quality of the purchased product.
In the future, the scientists plan to increase the accuracy of their model by expanding the dataset, testing other neural networks and ensemble methods, analyzing the temporal dynamics of bubbles (as opposed to static images) and combining their approach with physical and chemical methods to provide a fuller analysis.
The researchers also plan to study products not only from the food industry, but chemical products and pharmaceuticals, as well. For instance, bubble analysis can be helpful in quality control for liquid solutions, medical products that use cavitation to create nanoparticles or emulsions, and reactors with multiphase environments wherein gas-liquid processes depend on bubble dynamics.
