Planetary Prediction Engine Automatically Assembles Geospatial Data and Models, Achieving 83.3% Recall@10 for Ebola Hotspots

A Google research team has unveiled the Planetary Prediction Engine (PPE), an effort to delegate the most labor-intensive parts of geospatial prediction—finding data, verifying spatial and temporal coverage, combining features, and selecting models—to agents. After a user submits a natural-language objective and labeled data, the system first classifies the task as spatial regression, super-resolution, a diffusion model, or epidemic nowcasting. It then searches Data Commons, Google Earth Engine, government portals, and academic databases for covariates that can be joined to the data.

Rather than placing large data tables in the LLM context, PPE passes DataFrames, GeoJSON, and mobility matrices between three isolated stages using opaque handles. The LLM is responsible only for interpreting requests and selecting predefined tools. The data layer fuses statistical variables with frozen geospatial embeddings from PDFM and AlphaEarth, then searches across candidate architectures including linear models, boosting, XGBoost, and MLPs. A Feature Gate excludes components of the target formula, variables derived from the same survey source, data that would cause temporal leakage, and variables judged to be outcomes rather than causes. Missing-value statistics are also calculated exclusively from the training split.

The authors report that the full pipeline achieved an average R² of 76.8% across 21 U.S. CDC health indicators, compared with 60% for a manually constructed pipeline. For food-security downscaling in Nigeria, performance increased from 31.5% to 66.1%. Across five weekly forecasts during the 2026 Bundibugyo Ebola outbreak in the Democratic Republic of the Congo, the system captured all 15 newly infected districts among the 10 highest-risk districts for each period, producing an overall Recall@10 of 83.3%, versus approximately 73% for the consensus baseline.

This is not yet a general-purpose agent that can be deployed directly for public-health decision-making. The Ebola validation covers only one outbreak, and the filtering of causal direction has not yet been formalized. In one social-vulnerability downscaling test, high-resolution AlphaEarth features reduced R² from 52.0% to 40.1%, showing that additional satellite signals can also introduce spatial noise. The paper also does not currently provide a complete reproducible codebase. Key next steps include prospective validation across pathogens and countries, as well as publicly quantifying the cost of data and model search.