Google Research has developed the Planetary Prediction Engine (PPE), an experimental research capability within Google Earth AI, designed to automate the creation of high-fidelity geospatial models. The system addresses the challenge of fragmented data ecosystems and manual workflows that hinder rapid response during global crises. PPE’s architecture utilizes an LLM-orchestrated, three-stage process for data selection, curation, and model building, aiming to reduce model development time from weeks to minutes.
The PPE’s workflow begins with intelligent geospatial data selection, translating natural language queries into geographic constraints and retrieving relevant data from established repositories like Data Commons and Google Earth Engine, as well as performing open-web discovery. The second stage involves multimodal dataset curation, fusing data with embeddings from foundation models like Population Dynamics Foundation Models (PDFM) and AlphaEarth. A ‘Feature Gate’ enforces anti-leakage criteria to ensure data integrity. Finally, the system automatically builds and evaluates predictive models, utilizing families like regularized linear models and gradient-boosted decision trees, employing an ‘Overfitting Guard Protocol’ to prevent overfitting.
Evaluations across diverse benchmarks demonstrate significant improvements. In US public health and environment modeling, PPE achieved a mean R² of 76.8% for CDC health indicators, compared to a manual pipeline’s 60.0%. Similarly, in Nigeria’s food security analysis, PPE doubled accuracy when downscaling from provincial to local government levels. PPE also demonstrated effectiveness in real-time epidemiological nowcasting, improving Ebola outbreak predictions in the DRC.
The system’s modular design, with opaque handles for data transfer, avoids context-window limitations. The PPE’s autonomous operation and rapid insight generation represent a shift towards automated planetary analytics, particularly critical for humanitarian response.
Source: https://research.google/blog/planetary-prediction-engine-automating-global-models-via-earth-ai/