Google Automates Planet-Scale Prediction Workflows
Google’s experimental Earth AI agent discovers geospatial data, builds models and produces forecasts for health, food security and disaster risks.
From a question to a geospatial model
Google Research has introduced the Planetary Prediction Engine, an experimental agentic system that automates an end-to-end geospatial modeling workflow. Starting from a natural-language question, PPE identifies relevant signals, retrieves and cleans data, constructs features, trains candidate models, evaluates them and produces predictions with a report.
The system draws structured information from resources including Data Commons and Google Earth Engine, while searching government and academic sites when suitable variables are unavailable in established repositories. It combines those variables with embeddings from Google’s AlphaEarth and Population Dynamics foundation models. Automated gates are designed to reject future information, downstream effects and other features that could leak the target into the training data.
Google evaluated PPE across public health, environmental risk, food security and epidemiology. Across 21 US Centers for Disease Control and Prevention indicators, it reported a mean R² of 76.8%, compared with 60% for a manually constructed expert pipeline. For Nigerian food-security downscaling, PPE reached 66.1% against a 31.5% baseline. During a 2026 Ebola outbreak in the Democratic Republic of the Congo, it identified 15 of 18 newly affected health zones across five weekly forecasts, producing Recall@10 of 83.3%.
The supporting paper first appeared on arXiv before the edition window, but Google publicly introduced the experimental capability during this edition. PPE is not presented as a generally available product, and the reported comparisons come from its developers rather than independent deployment studies.
Why it matters
Geospatial forecasting often takes specialized teams weeks because useful data is scattered across incompatible sources and must be aligned by location and time. Automating that work could let public-health and humanitarian teams build provisional models in minutes during fast-moving crises. The risk is that speed may obscure fragile proxies, uneven regional data and false confidence. PPE’s value will ultimately depend on transparent provenance and expert review, especially when its maps influence where scarce aid or medical resources are sent.