Planetary prediction engine: Automating global models via Earth AI

Details

Google Research has introduced the Planetary Prediction Engine (PPE) as an experimental capability under its Earth AI initiative, targeting a real bottleneck in global modeling: the weeks-long manual process of data curation, feature engineering, and spatial validation that specialized teams must complete before any predictive model can be trained. PPE automates this entire workflow from a single natural-language query. The system operates in three modular stages, each orchestrated by an LLM. Stage 1 handles intelligent geospatial data selection: it translates a natural-language query into geographic constraints (spatial granularity, temporal scope, join keys), then conducts 'grounded signal discovery' — forming domain hypotheses and identifying both direct and proxy signals cross-referenced against published scientific literature. Details of stages 2 and 3 are available in the accompanying paper. PPE's target use cases are explicitly humanitarian: food security forecasting, disease outbreak tracking, environmental disaster risk mapping, and socioeconomic vulnerability analysis. Benchmark results show improvements over baselines across these domains without manual intervention. The authors emphasize the speed advantage — from weeks to minutes — as critical for rapid response during humanitarian crises. This is experimental research, not a productized Google Cloud offering. A companion paper has been published, making it accessible for the research community to reproduce and build on.