Planetary Prediction Engine

Planetary Prediction Engine (PPE) is an experimental geospatial modeling system developed by Google Research. It takes a natural-language prediction request, finds and prepares geographic data, adds pretrained spatial embeddings, searches among several model families, and returns predictions with a report. Google introduced PPE on August 27, 2026 as an early-stage research capability within Google Earth AI, its broader program for applying artificial intelligence to planetary data.^[1][2]

PPE is not a single global forecasting model. It is an agentic workflow for building a separate model for a specified task, place, time period, and spatial resolution. The research prototype was evaluated on public-health and environmental regression in the United States, food-security downscaling in Nigeria, and outbreak-hotspot prediction in the Democratic Republic of the Congo. The reported results come from a Google-authored preprint and announcement; they had not been independently reproduced as of August 2026.^[1][2]

Overview

PPE belongs to the field of GeoAI, where machine-learning methods are applied to spatial data and geographic problems. Its stated purpose is to automate work that usually requires separate rounds of domain research, data discovery, geographic joins, feature engineering, model selection, and spatial validation. The paper describes the input as a labeled dataset plus a natural-language request. The output is an executed model, predictions, and a report rather than a general-purpose conversational answer.^[2]

The system first classifies the request as spatial regression, super-resolution downscaling, spatial transmission, or epidemiological nowcasting. That classification constrains later choices. An outbreak task can trigger construction of mobility and travel-time features, while ordinary spatial regression can focus on static demographic and environmental variables. Off-the-shelf large language models act as black-box orchestrators that interpret the request and choose from a fixed set of tools. The paper does not identify the LLM versions used.^[2]

Scroll sideways for more →

Workflow

Intelligent data selection

The first stage translates the request and supplied labels into geographic and temporal constraints. It determines the spatial granularity, join-key format, area of interest, and prediction window. It then proposes direct signals and causal proxy signals, checks those hypotheses against literature or official reports, and looks for corresponding datasets.^[1][2]

The retrieval process starts with established sources. Data Commons supplies demographic and economic statistics, while Google Earth Engine supplies raster and Earth-observation layers. The paper also lists Google Maps Platform Insights for aggregated point-of-interest statistics, OpenStreetMap and the Humanitarian Data Exchange for roads and boundaries, and Open-Meteo and NOAA for weather variables. If those sources do not cover a proposed signal, PPE can search government and academic repositories on the open web for files such as CSV, GeoJSON, and Parquet datasets.^[2][6][7]

Candidate datasets receive scores for provenance and licensing, spatial and temporal fit, signal alignment, format quality, and redundancy. Provenance and license openness carry the highest weight in the paper's rubric. Retained data are standardized to common geometry, date, and join-key conventions and merged into a covariate table. The system also produces source-audit and provenance metadata.^[2]

Multimodal dataset curation

The second stage combines explicit statistical or environmental variables with pretrained geospatial embeddings. The Population Dynamics Foundation Model provides 330- or 512-dimensional representations intended to encode population, demographic, mobility, health, and economic patterns at administrative boundaries. AlphaEarth Foundations supplies 64-dimensional satellite-derived embeddings representing features such as land cover, vegetation, topography, and land use.^[2][4][5]

PPE uses these models as frozen feature extractors. It does not fine-tune them for each downstream task. Tabular covariates can receive skew correction, percentile clipping, collinearity filtering, and standard scaling. Embedding vectors receive L2 normalization, but the system avoids independently standardizing their dimensions because that could alter relationships in the learned representation space.^[2]

The data inventory mixes sources with different access conditions. AlphaEarth's annual satellite embedding dataset is available through Earth Engine. The PPE paper labels PDFM as pre-general-availability and available for research use. Google Maps Platform Insights can incur commercial cloud billing. These conditions matter because access to an input dataset or embedding does not provide access to PPE itself.^[2][5]

Model search and prediction

The final stage trains and evaluates models on the curated data. Its search space includes Ridge, Lasso, and Elastic Net regularized linear models; histogram-based gradient boosting; XGBoost; and multilayer perceptrons. The stage can use a fixed-seed random 80/20 split, a geographic group split, or three-fold cross-validation, depending on the task and sample size. Reported metrics include R-squared, root mean squared error, mean absolute error, mean absolute percentage error, and Pearson correlation. Outbreak-hotspot experiments also use Recall@10.^[2]

PPE's modular design keeps the prediction stage from retrieving or modifying data. Curated DataFrames, GeoJSON geometries, and mobility matrices pass between stages through opaque handles instead of being serialized into LLM prompts. According to the paper, this avoids filling the orchestrating model's context window with large data artifacts and limits each stage to defined inputs and outputs.^[2]

Leakage and overfitting safeguards

The paper describes a Feature Gate that evaluates every proposed covariate before modeling. It rejects four kinds of information: mathematical components or subindices used to construct the target; variables derived from the same survey or imputation source as the target; variables that are causally downstream of the target; and observations from after the prediction window. For population-related targets, the protocol also favors intensive rates over enumerative counts to reduce direct census-count leakage.^[1][2]

Missing data are handled with split-isolated imputation. PPE computes imputation statistics only from the training partition, then applies them to validation and test partitions. Features with inadequate coverage are dropped. For spatial evaluation, the system can group the split geographically so that nearby observations do not leak local structure across training and evaluation sets.^[2]

An Overfitting Guard assesses sample size, the feature-to-sample ratio, and spatial grouping before training. Medium-risk cases restrict tree depth, while high-risk cases use stronger regularization and favor regularized linear models. If validation produces a negative score or a large train-validation gap, PPE can discard the candidate and run one more search with more conservative settings. The one-iteration limit is intended to prevent an open-ended optimization loop.^[2]

These procedures reduce known leakage risks but do not establish that leakage is impossible. The authors identify formal verification of causal-direction filtering as an open problem, especially when a variable and target influence each other. Several United States spatial-regression benchmarks also follow conventional random train-test splits used by prior work, rather than the stricter question of generalization to wholly unseen regions.^[2]

Evaluation

All results in this section were reported by PPE's authors. The evaluations use different targets, splits, and metrics, so their scores should not be combined into one measure of system quality.^[1][2]

DRC outbreak hotspot prediction

The outbreak experiment used records from the 2026 Bundibugyo ebolavirus outbreak in the Democratic Republic of the Congo. Data covered 519 health zones and included INRB case records, WHO reports, population and environmental layers, roads, buildings, health facilities, travel times, and mobility estimates. PPE made five sequential one-week forecasts. Recall@10 measured how many newly invaded health zones appeared among the ten zones ranked at highest risk.^[2][10]

The authors report that full PPE identified 15 of 18 newly invaded zones, for Recall@10 of 83.3% with a 95% confidence interval of 60.8-94.2. A covariate configuration reached 77.8% (54.8-91.0), and the cited Bayesian baseline was approximately 73%. The reported improvement over that baseline is 10.3 percentage points.^[1][2]

An appendix compares two outbreak objectives. Spatial transmission regression reported test Top-10 accuracy of 0.8 for ranking vulnerable zones, while a mechanistic SEIR configuration reported test RMSE of 0.9204 for case-volume tracking. These metrics evaluate different outcomes and do not show that either approach is universally better. The entire PPE outbreak evaluation covers one event and is retrospective; the paper does not report a prospective public-health deployment.^[2]

Nigeria food-security downscaling

The Nigeria benchmark trained on monthly Food Consumption Group prevalence for 30 state-level units and evaluated predictions for 581 local-government areas across 40 months. Features included time trends, food-price indicators, vegetation measures, rainfall, nighttime lights, PDFM embeddings, and AlphaEarth embeddings. The paper says the WFP household records were anonymized and contained no personally identifiable information. Local-government-area estimates produced through multilevel regression and poststratification were kept blind during training.^[2]

Scroll sideways for more →

The authors also report out-of-fold mean absolute error of 10.0% at the finer resolution, compared with 13.6% for the baseline. Google's description that PPE "doubled baseline accuracy" refers to the R-squared values, 66.1% versus 31.5%; it does not mean the error rate was cut in half.^[1][2]

United States downscaling and spatial regression

For county-to-ZCTA Social Vulnerability Index downscaling, the detailed results report R-squared of 11.0% for a macro-covariate and interpolation baseline, 25.6% for geospatial covariates, 36.9% for PDFM alone, and 37.6% for covariates plus PDFM. A separate ablation discussed by the authors found that adding high-resolution AlphaEarth features to a covariate-plus-PDFM configuration reduced R-squared from 52.0% to 40.1%. They interpret that result as a possible mismatch between high-frequency physical features and a socioeconomic target at a different administrative scale, not as a general failure of satellite embeddings.^[2]

The authors report the following aggregate spatial-regression results:

Scroll sideways for more →

The CDC result is 16.8 percentage points above 60.0%. One sentence in the preprint calls it a 23-point improvement, which does not match the values in its table and is not repeated here. The SVI headline and detailed table use differently labeled baselines, so 58.6% and 60.3% should not be treated as the same number.^[1][2][8][9]

Ablation findings

The paper organizes experiments into baseline, covariate-only, embedding-fusion, and full-stack tiers. In the DRC hotspot task, adding geospatial covariates raised Recall@10 from approximately 73% to 77.8%, while the complete feature-selection and embedding pipeline reached 83.3%. In Nigeria, vegetation contributed more than nighttime lights when each was added to the temporal baseline, and the automated full selection produced the highest reported R-squared.^[2]

For the 20 FEMA targets, adding Data Commons variables to PDFM and AlphaEarth produced a nationwide mean R-squared of 61.1%, compared with 59.9% for the manual expert configuration. Adding the full Intelligent Data Selection pipeline raised the reported mean to 64.9%. Category results varied: full PPE reported 69.4% for socioeconomic and composite targets, 68.3% for atmospheric and climatological targets, and 56.2% for geophysical and hydrological targets. The variation, along with the SVI downscaling regression when AlphaEarth was added, shows that more feature sources did not improve every target automatically.^[2]

Availability and relationship to Earth AI

Google Earth AI is the broader initiative that groups Google's geospatial models, datasets, reasoning systems, and related products. PPE is one experimental modeling capability within that initiative. It uses Earth AI resources such as Earth Engine, AlphaEarth, and PDFM, but it is not another name for Google Earth AI.^[1][3]

As of August 28, 2026, Google's announcement and the paper did not provide a public PPE endpoint, source-code repository, downloadable implementation, model weights, pricing, or access request process. The Google article links to the preprint and describes PPE as early-stage research. AlphaEarth embeddings can be accessed separately through Earth Engine, while the paper describes PDFM access as pre-general-availability research use. Neither form of input access amounts to availability of the complete PPE system.^[1][2][5]

Limitations

The study is a first-party evaluation reported in a preprint. It does not provide independent reproduction, peer-reviewed confirmation, or evidence of sustained production use. "Autonomous" refers to the tested workflow running after a natural-language specification under the authors' setup; it does not mean that data selection, causal judgments, or outputs are guaranteed correct.^[1][2]

The authors list four main technical limitations. PDFM and AlphaEarth remain frozen rather than tuned for the downstream tasks. High-resolution satellite features reduced performance in one SVI downscaling configuration. Causal-direction filters lack formal verification. The outbreak evaluation covers only one pathogen event, geography, and surveillance environment. They identify longer forecast horizons, streaming sources, and combinations of transmission and mechanistic outbreak models as future work.^[2]

The benchmark matrix also mixes random splits, geographic splits, cross-validation, coarse-to-fine evaluation, and rolling forecasts. Each design tests a different form of generalization. Results from United States census-tract regression do not establish performance in data-scarce regions, and retrospective hotspot ranking does not establish prospective clinical or emergency-response effectiveness.^[2]

References

1. ^Rama Pasumarthi and Shravya Shetty. "Planetary prediction engine: Automating global models via Earth AI." Google Research, August 27, 2026. research.google/...ting-global-models-via-earth-ai
2. ^Evelyn Ma, Rama Kumar Pasumarthi, Kishwar Shafin, et al. "Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings." arXiv:2608.26088v1, August 26, 2026. arxiv.org/...2608.26088
3. ^Google. "Google Earth AI." Accessed August 28, 2026. ai.google/earth-ai
4. ^Mayank Agarwal, Mimi Sun, Chandler Kamath, et al. "General Geospatial Inference with a Population Dynamics Foundation Model." arXiv:2411.07207, 2024. arxiv.org/...2411.07207
5. ^Christopher F. Brown, Michal R. Kazmierski, Valerie J. Pasquarella, et al. "AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data." arXiv:2507.22291, 2025. arxiv.org/...2507.22291
6. ^Data Commons. "Data Commons documentation." Accessed August 28, 2026. docs.datacommons.org
7. ^Google Earth Engine. "Earth Engine Data Catalog." Accessed August 28, 2026. developers.google.com/...datasets
8. ^Centers for Disease Control and Prevention. "PLACES: Local Data for Better Health." Accessed August 28, 2026. cdc.gov/places
9. ^Federal Emergency Management Agency. "National Risk Index." Accessed August 28, 2026. hazards.fema.gov/nri
10. ^World Health Organization. "Ebola disease caused by Bundibugyo virus - Democratic Republic of the Congo." Disease Outbreak News, August 1, 2026; revised August 3, 2026. who.int/...2026-DON614