Google Research has introduced an experimental planetary prediction engine that builds geospatial machine-learning models from natural-language questions. The system handles the workflow from finding and cleaning data through feature engineering, training, and evaluation.
Geospatial projects often combine satellite imagery, maps, weather records, population data, and other sources with different formats and coverage. Google says specialist teams can spend weeks preparing those inputs before modeling begins. The new engine is designed to search Earth AI’s planetary data assets and assemble a prediction pipeline without starting from a pre-curated table.
Researchers tested the approach across public health, food security, environmental risk, and socioeconomic prediction tasks. The system selects relevant data, creates spatial and temporal features, compares models, and returns evaluation results. Google reports improvements across diverse tasks, but describes the capability as experimental rather than a production forecasting service.
Automation could shorten the path from a question to an initial model, especially during humanitarian responses. Its output still depends on data quality, geographic coverage, and suitable validation. High-stakes decisions would require domain experts to inspect assumptions, check regional bias, and determine whether the model remains reliable outside its evaluation setting.