Google Maps uses DeepMind GNNs to predict traffic before it happens

Google Maps has upgraded its Estimated Time of Arrival (ETA) technology by integrating Graph Neural Networks (GNNs) developed in partnership with DeepMind. This shift allows the navigation giant to move beyond reactive data, predicting traffic congestion that may not have materialized by the time a driver begins their trip.

The system generates predictions by blending live aggregate location data from millions of users with historical traffic patterns for specific road segments. While Google reported that its ETA predictions were already accurate for over 97% of trips before the GNN enhancements, the new architecture allows the AI to analyze road networks as a graph rather than a series of isolated segments. This technique enables the app to better predict whether a user will be affected by a slowdown that has not yet started.

The shift to agile modeling

For more than a decade, Google Maps relied primarily on historical data and real-time user pings to estimate travel times. However, the complexity of modern urban traffic and sudden global disruptions—most notably the 2020 COVID-19 lockdowns—exposed the limitations of static historical models. To maintain accuracy during these shifts in global mobility, Google updated its models to prioritize more recent historical data, specifically focusing on patterns from the previous two to four weeks.

This evolution represents a fundamental transition from simple linear predictions to a holistic understanding of how traffic flows through a connected network. By utilizing GNNs, the system can better understand the spatial relationships between different road segments, recognizing how a bottleneck in one area will inevitably ripple through connected streets.

Implications for urban mobility

Improving the precision of ETA predictions has consequences that extend beyond user convenience. Accurate timing is a critical component of urban logistics and the efficiency of time-sensitive travel. By anticipating future congestion, Google can proactively reroute millions of drivers before they enter a traffic jam, which potentially reduces overall city congestion and lowers fuel consumption across the network.

The road ahead

As Google continues to refine its partnership with DeepMind, the focus remains on increasing the agility of these AI models to handle unpredictable urban environments. While the current integration of GNNs has significantly enhanced spatial awareness, the company continues to balance the use of long-term historical trends with short-term data bursts to ensure the system remains resilient to sudden changes in driver behavior.