Google Research Shows an Experimental System That Builds Geospatial Models From a Plain-Language Query

Google Research introduced the Planetary Prediction Engine (PPE) on August 27, 2026 as part of Google Earth AI. It is described as an experimental research capability that takes a natural-language query and autonomously runs the workflow from data discovery through model training, evaluation and report generation¹.

• The pipeline is split into three LLM-orchestrated stages: (1) geospatial data selection, (2) multimodal dataset curation, and (3) model building and prediction¹
• Data is retrieved from Data Commons and Google Earth Engine, and fused with embeddings from Population Dynamics Foundation Models (PDFM) for socio-demographics and AlphaEarth for satellite imagery¹
• Reported results: a mean R² of 76.8% across 21 CDC health indicators in the US (60.0% for a manual expert pipeline), R² 66.1% versus 31.5% when downscaling food security in Nigeria, and Recall@10 of 83.3% for nowcasting an Ebola outbreak in the DRC (against a published Bayesian baseline of about 73%)¹
• Google says this cuts the time to build complex prediction models from weeks to minutes¹

All of these figures are Google’s own reporting, not third-party verification. The company describes PPE as an early-stage research project, and there is no announcement of general availability¹. That said, if you are weighing how far to let a model drive an analysis pipeline, the gaps against the manual baselines here are a concrete data point on what happens when data discovery and feature engineering are handed over too.

Sources

1. Planetary prediction engine: Automating global models via Earth AI - Google Research official blog (August 27, 2026)