Researchers at the University of Adelaide have developed machine learning models that predict wine quality based on vineyard conditions during the growing season, potentially helping winemakers optimise harvest timing and vineyard management practices.

The system analyses soil moisture, temperature patterns, vine canopy characteristics captured by drones, and other environmental data collected throughout the growing season. By March or April, months before harvest, it can predict whether a vintage will be exceptional, average, or disappointing with about 80% accuracy.

That early warning gives winemakers time to adjust plans. A predicted poor vintage might lead to more aggressive crop thinning to concentrate quality in remaining fruit. A predicted exceptional vintage might justify investments in premium oak barrels or extended aging.

Professor Mark Krstic, who leads Adelaide’s wine research program, said the models work because wine quality correlates strongly with specific weather patterns and vine stress during critical growth stages. “It’s not just about total rainfall or average temperature. What matters is the timing of heat events, water stress during fruit development, and diurnal temperature variation during ripening.”

Winemakers have always understood these relationships, but human assessment is qualitative and relies on experience. The machine learning approach quantifies these relationships and can integrate far more data than humans can mentally process.

The research team trained their models on fifteen years of data from Barossa Valley vineyards, including detailed environmental monitoring and subsequent wine quality assessments by expert tasters. That historical data provides the ground truth needed for supervised machine learning.

The models use gradient boosted decision trees, an ensemble machine learning method that performs well on heterogeneous data with complex interactions. Neural networks were tested but didn’t perform better, probably because the training dataset wasn’t large enough to support deep learning.

One interesting finding is that vine water stress timing matters more than total stress. Moderate water stress during veraison, when berries begin ripening and changing colour, tends to improve quality by concentrating flavours. But stress during earlier growth stages reduces yields without quality benefits.

The models capture these non-linear relationships automatically from data rather than requiring researchers to specify them explicitly. That’s valuable because some quality determinants aren’t well understood mechanistically.

Early predictions aren’t perfect. The models occasionally predict poor vintages that turn out well, or vice versa. Forecast accuracy improves as harvest approaches and more data becomes available, but that reduces the time available for management adjustments.

The research team is now working with several South Australian wineries to test the system commercially. Participating wineries get early vintage predictions while researchers gain access to additional data for model refinement.

One challenge is adapting models trained on Barossa Valley data to other wine regions. Climate, soil types, and grape varieties differ between regions, meaning models need retraining or adaptation. Transfer learning techniques might help, using models trained on one region as starting points for another region with limited data.

The system could eventually become a commercial service where wine regions collectively fund monitoring infrastructure and model development, then individual wineries subscribe to access predictions. Several agricultural technology companies have expressed interest in licensing the technology.

Economic value is substantial if predictions are accurate. A 5% yield reduction applied at the right time to improve quality could increase wine value by 20-30% for premium wines. Conversely, knowing early that a vintage will be disappointing lets wineries adjust marketing and pricing expectations.

The research builds on decades of viticulture and oenology research at Adelaide, which has one of the world’s leading wine science programs. Australian wine research generally is strong, reflecting the industry’s economic importance and export orientation.

Australia exports over $2 billion worth of wine annually, making it a significant agricultural sector. Climate change poses challenges through changing weather patterns, earlier harvest dates, and shifting suitability of traditional wine regions.

Predictive models that help winemakers adapt to varying conditions could become increasingly valuable as climate variability increases. The research team is specifically investigating whether AI systems can help identify adaptation strategies like variety selection, irrigation scheduling, or canopy management.

Some wine traditionalists are skeptical about AI’s role in winemaking. They argue wine quality is fundamentally about craftsmanship and terroir, not algorithmic optimisation. The counter-argument is that AI provides tools to enhance human decision-making, not replace it.

The Adelaide models don’t make winemaking decisions; they provide information that winemakers interpret based on their goals and expertise. Think of it as advanced weather forecasting for wine, useful but not deterministic.

Data collection remains challenging for smaller wineries that lack resources for extensive monitoring. The research team is exploring whether satellite and weather station data can partially substitute for in-vineyard sensors, making the system accessible to more producers.

Privacy and data ownership are also considerations. Wineries want quality predictions but may be reluctant to share detailed vineyard data that could reveal proprietary practices or assist competitors. The research team is developing privacy-preserving approaches where models train on aggregated or anonymised data.

The wine quality prediction project is part of broader efforts to apply AI in Australian agriculture. Similar approaches are being developed for grain quality prediction, livestock management, and pest outbreak forecasting.

Whether AI-assisted viticulture achieves widespread adoption depends partly on proving economic value through demonstrated quality and yield improvements. The Adelaide team expects to have multi-year results from commercial trials by 2027, which should clarify whether the approach delivers on its promise.

For now, it’s an interesting example of how machine learning can augment traditional expertise in industries like winemaking where quality assessment has historically been entirely human judgment-based.