Will AI in the future be able to analyse wine aromas and recognise faults and flavour types?
Prof Dominik Durner from the "PINOT" project at the Neustadt Wine Campus told Kristine Bäder what the technology is currently capable of - and what our nose can do better.
Prof Dominik Durner heads the PINOT project.
Stephan PresserIs it possible to analyse flavours with the help of AI via sensors and translate them into human language? Scientists and researchers from various disciplines have been working on this question for three years as part of the PINOT project funded by the German Federal Ministry of Agriculture. Their vision: a flavour detector for professionals in the cellar and in retail. Prof Dr Dominik Durner from the Neustadt Wine Campus draws a positive interim balance.
The start-up Genie Enterprise, based in Ludwigshafen, Germany, approached the Professor of Oenology back in 2019 with the idea of developing a "sommelier made of AI". The founders were open to the question of whether the capabilities of AI could also be transferred to the world of flavours. They quickly found academic partners in Trier University of Applied Sciences and the Fraunhofer Institute in Erlangen. With Genie Enterprise, Wille Engineering and the start-up Vineyard Cloud, three companies supported the project to develop a practical application of the research results. The research project was ultimately funded by the Ministry of Agriculture via the "AI in Agriculture" funding programme.
The first task was to interconnect electrochemical sensors to test which odour signals the sensors recognise - and whether they are comparable to the human nose. "The first tests were disastrous and sobering, they didn't work at all," recalls the oenologist. "The machines were simply paralysed by the ethanol." So the research group began to train the sensors for wine faults, which "worked more or less well." Durner explains how the team developed the project: For example, a sensor is first given 100 wines with a clear bockser, or hydrogen sulphide. It is then presented with the first wine without this ingredient and has to recognise the difference. This flavour perception is then validated by the researchers. "This is real Sisyphean work," emphasises Dominik Durner. Compounds such as hydrogen sulphide are still relatively easy to learn. "However, the recognition of complex flavours can only be achieved with big data. 100 wines are simply nothing." Instead, the sensors have to be fed with data from thousands of wines. For every grape variety and every wine defect.
Even if the capabilities of the artificial nose are still limited, Prof Durner sees enormous potential for applications after three years of research - even beyond wine. "The performance of a cellar master is not the same every day - that of a digital system always works the same and is not affected by emotions," he explains the possible integration into the production process in the cellar. Currently, the machine can do "less than 0.1 per cent of what the human nose can do."
However, when it is properly trained, its capabilities are enormous. At the current stage, for example, it reliably recognises a bump. "This already works at a time when the human nose is still a long way from recognising hydrogen sulphide," says the professor, explaining the successes. Winemakers could react very early in the cellar during production if the fermentation process is not running smoothly. AI could also become interesting for the trade when it comes to the authenticity of wines. With the help of AI systems, it should be possible for in the future to store a kind of fingerprint of the wine and compare it later. This would allow buyers to ensure that the wine delivered corresponds to the sample ordered. However, AI systems are unlikely to make a career as judges in tasting competitions for the time being. "Teaching AI the wine profile of a Mosel Riesling, for example, is much more difficult," Dominik Durner concedes. The same applies to the qualitative assessment of a wine, as professional tasters can do.
This already fails due to the lack of databases. With the start of the follow-up project "WineIO", which is also funded by the Ministry of Agriculture, research will continue on the topic for a further three years. Durner has also contacted the University of Davis in California because there are wineries there with a corresponding "love of data". "In the United States, people are much more data-driven than here. A lot is already measured and documented there for forecasting and development," he says, hoping for positive input. His aim is to utilise the results of his research commercially within winemaking. A device, ideally in mobile phone format, that is used in the cellar during fermentation would have the advantage of being able to react directly and not having to wait for analysis results from the laboratory.
The future AI-controlled technology could also be transferable to other areas outside the wine industry: "Such a device can be used wherever volatile substances are involved." For example, as a replacement or supplement for drug and explosives detection dogs or in medicine. "It wouldn't be the first time that research from the wine sector has been successfully utilised there," explains Durner. He does not see the danger of wines tasting uniform as a result of AI in the future: "On the contrary. Artificial intelligence offers the opportunity for targeted production. Consumer orientation is also important in the wine industry."