IN Brief:
- The project analyses volatile chemical signatures emitted by grapevines, juice and wine using sensing and machine learning.
- Researchers identified samples at elevated risk of Fresh Mushroom Aroma before the defect would normally become detectable.
- Further work will examine applications including vine disease, soil conditions and quality effects associated with wildfires.
Analog Devices, Moët Hennessy and the University of California, Davis are developing a chemical sensing platform that uses machine learning to identify volatile signatures associated with emerging quality risks in grapes and wine.
The first reported demonstration focused on Fresh Mushroom Aroma, a defect that can develop during wine production. Researchers at Moët Hennessy’s Robert-Jean de Vogüé Research Center built a library of samples and associated data, then used it to train algorithms to recognise chemical patterns linked with an increased risk of the defect. The system identified samples at elevated risk before Fresh Mushroom Aroma would normally have been detected through conventional methods.
The sensing task is broader than measuring one known compound against a fixed threshold. Grapevines, juice and wine emit combinations of volatile chemicals that change with biological processes and environmental conditions. The platform captures those signatures and uses the pattern across the measurements as the input to a machine learning model. Its output depends on the relationship established between the measured pattern and reference samples whose condition is already known.
That architecture places equal weight on the sensor and the dataset used to interpret it. A sensitive instrument can detect changes that have little diagnostic value, while an algorithm cannot compensate for measurements that drift unpredictably between samples. Repeatability, calibration, sample handling and environmental variation therefore determine whether a model trained in one study continues to recognise the same chemical condition when the location, vintage or operating conditions change.
Analog Devices has not disclosed the detailed sensing element, sampling arrangement or complete signal chain used in the current programme. The company says the platform identified complex chemical signatures associated with elevated Fresh Mushroom Aroma risk in the samples tested, but it has not published limits of detection, false positive rates or comparative accuracy against established laboratory methods. Those measurements will be needed before the platform can be assessed as a routine production tool rather than a research system.
Moët Hennessy contributes the sample library and winemaking expertise, while UC Davis adds viticulture and enology research. Analog Devices is developing the underlying sensing platform from technology that originated in earlier research at the Massachusetts Institute of Technology. The partners are also examining whether the same approach can identify vine disease, assess soil conditions and detect quality effects associated with environmental events including wildfires.
Each new application will require its own reference data. A model trained to recognise the chemical signature associated with Fresh Mushroom Aroma cannot be assumed to identify wildfire taint or disease without examples that connect those conditions with the sensor output. Building the dataset is therefore part of the engineering programme rather than a single step completed when the sensing hardware is installed.
Earlier detection could change how a winery uses analytical testing. Conventional laboratory methods remain valuable when the compound or defect to be measured is already understood and an established analytical procedure exists. A system based on chemical signatures may be more useful when several volatile compounds change together or when the relevant pattern appears before the defect becomes obvious through sensory assessment or a targeted test.
The same measurement principle could extend beyond wine because fermentation, agriculture, food production and other biological processes generate complex volatile mixtures whose composition changes over time. A sensor that captures a reproducible chemical fingerprint, combined with labelled data linking that fingerprint to a process condition, could provide an earlier indication of change without requiring every condition to be reduced to one marker compound.
The current project demonstrated that the measured signatures contained enough information for the trained model to identify elevated Fresh Mushroom Aroma risk in the samples tested. It has not yet established a commercial instrument specification or deployment timetable. Repeatability across larger datasets and operating conditions will determine whether the relationship between the chemical signal and the underlying quality risk remains stable outside the original study.
