Google Earth AI has introduced the planetary prediction engine (PPE), an experimental system that autonomously executes end-to-end geospatial prediction workflows from natural-language queries. PPE discovers and validates candidate signals, fuses structured covariates with foundation-model embeddings, and automatically trains and evaluates predictive models, producing reports without manual intervention.
Why this matters
Addressing global challenges — from regional food security and environmental hazards to real-time disease monitoring — requires high-fidelity geospatial models. Building these models traditionally demands specialized teams and weeks of manual data curation and feature engineering. PPE aims to shorten that timeline and lower the need for engineering resources by automating the full pipeline.
How the PPE is structured and operates
PPE divides the predictive workflow into three modular stages, each coordinated by a large language model (LLM).
1) Intelligent geospatial data selection
Given a natural-language query, PPE translates intent into strict geographic and temporal constraints (spatial granularity, join-keys, temporal scope). It then performs grounded signal discovery: formulating domain hypotheses and identifying direct and causal proxy signals validated against published literature. The system retrieves covariates from established geospatial repositories such as Data Commons and Google Earth Engine. For signals not present in those repositories, it conducts live open-web discovery of government portals and academic repositories at inference time.
2) Multimodal dataset curation
Collected covariates are fused with pre-trained geospatial foundation model embeddings — Population Dynamics Foundation Models (PDFM) for socio-demographic latent states and AlphaEarth for satellite imagery semantics. PPE enforces a strict "Feature Gate" that mitigates target leakage by evaluating every candidate covariate against four anti-leakage criteria (filtering out mathematical subcomponents, shared survey data, downstream causal effects, and future temporal data) to preserve evaluation integrity.
3) Automated model building and prediction
PPE searches over multiple model families (regularized linear models, gradient-boosted decision trees — GBDT, and multi-layer perceptrons) and applies an Overfitting Guard Protocol that pre-assesses dataset risk and implements a self-correction loop to detect and recover from generalization failures. Each stage operates with well-defined inputs and outputs; data artifacts are passed between stages via opaque handles instead of serializing them into LLM prompts to avoid context-window limitations.
Performance across tasks and benchmarks
PPE was evaluated across a matrix of machine-learning paradigms, geographies, and scientific domains.
Spatial regression: US public health and environment
Across 21 CDC (Centers for Disease Control and Prevention) health indicators, PPE’s intelligent data selection and multimodal fusion achieved a mean R² of 76.8% compared to 60.0% for a manual expert pipeline. Similar gains were observed for FEMA (Federal Emergency Management Agency) national risk indicators (mean R² 64.9% vs. 60.0% baseline) and the Social Vulnerability Index (mean R² 66.2% vs. 58.6% baseline).
Super-resolution downscaling: Nigeria food security
In data-scarce regions, coarse regional reporting can hide local vulnerability. By autonomously integrating localized market shocks, food price anomalies, and microclimate indicators, PPE doubled baseline accuracy when downscaling food security from the provincial (ADM1) level to the local government area (ADM2) level (R² 66.1% vs. 31.5%), producing higher-fidelity vulnerability maps for humanitarian actors.
Epidemiological nowcasting: DRC Ebola outbreak
For real-time prediction of new transmission hotspots during the 2026 Bundibugyo ebolavirus outbreak in the Democratic Republic of the Congo, PPE achieved Recall@10 = 83.3%, correctly identifying 15 of 18 newly invaded health zones across five sequential weekly forecasts. This is an absolute improvement of 10.3 percentage points over the published state-of-the-art Bayesian modeling baseline (~73%), driven by fusing epidemiological signals with PDFM embeddings and PPE-selected geospatial covariates.
The role of multimodal fusion and intelligent selection
Across experiments, combining structured statistical covariates with latent foundation-model embeddings proved synergistic. Statistical covariates supply explicit, interpretable signals, while Population Dynamics Embeddings and AlphaEarth Foundation embeddings capture complex non-linear patterns learned during large-scale pre-training. Ablation studies show the multimodal fusion plus intelligent data selection consistently outperforms single-modality baselines.
Limitations and future directions
PPE is an early-stage research project. The authors plan to expand source coverage and incorporate additional foundation-model embeddings such as Remote Sensing Foundations multimodal embeddings. Operational deployment will need careful consideration of ethical, legal, and data-protection issues, particularly around sensitive or personal data.
Conclusion
The PPE demonstrates that autonomous AI can match or improve performance across diverse geospatial prediction tasks: estimating chronic disease prevalence in the United States, improving resolution of food insecurity in Nigeria, and nowcasting active viral outbreaks in the Democratic Republic of the Congo. By automating data discovery, multimodal fusion, and model optimization, PPE reduces the technical barrier to planetary-scale analytics and enables faster deployment in time-sensitive situations.
Acknowledgements
The authors acknowledge contributions from all paper co-authors and thank the UN World Food Programme (WFP) Vulnerability Analysis and Mapping (VAM) team for data and research support, the Institut National de Recherche Biomédicale (INRB) for collaboration on DRC Ebola nowcasting, and the teams behind Data Commons, Google Earth Engine, Population Dynamics Foundation Models, and AlphaEarth for providing foundational data and model infrastructure.
