行星预测引擎：通过Earth AI自动化全球模型

Rama Pasumarthi, Staff Software Engineer, and Shravya Shetty, Distinguished Engineer, Google Research As part of Google Earth AI , we introduce the planetary prediction engine (PPE), an experimental research capability that autonomously executes the full geospatial modeling workflow — from data discovery to model training — achieving improvements across diverse prediction tasks in public health, food security, environmental risk, and socioeconomics. Addressing humanity's most pressing global challenges — from forecasting regional food security and environmental disaster risks to tracking real-time disease outbreaks and mapping socio-economic vulnerability — requires high-fidelity geospatial modeling. However, building these models is hindered by a fragmented data ecosystem that requires specialized teams to spend weeks on manual data curation, feature engineering, and specialized spatial validation. While existing AutoML and LLM-based agents effectively automate standard machine learning pipelines, they rely on pre-curated tabular data and lack the specialized capabilities needed to autonomously handle geospatial workflows. Consequently, planetary-scale analytics remains a significant bottleneck, a limitation that is especially severe when rapid response is critical during humanitarian crises. Today we introduce the planetary prediction engine (PPE), the latest experimental research capability within our broader Google Earth AI initiative that turns planetary information into actionable insights. Previously, we demonstrated Earth AI’s reasoning capabilities in intelligently bringing together diverse geospatial assets. As an autonomous AI system, PPE executes the full geospatial prediction workflow, from planetary-scale data discovery and cleanup to training and evaluation, directly from natural-language queries. Given a geospatial predictive query, PPE will autonomously retrieve relevant data, engineer features, train and evaluate predictive models, and generate a comprehensive report. We demonstrate that by delivering improvements over diverse benchmarks without manual intervention, PPE effectively reduces the time needed to build complex planetary prediction models from weeks that include manual data engineering to mere minutes that result in autonomous insight. The PPE decomposes the predictive workflow into three modular stages, each orchestrated by an LLM: Stage 1: Intelligent geospatial data selection. Given a natural-language query, the PPE translates the prompt into strict geographic constraints (spatial granularity, join-keys, temporal scope). It then conducts grounded signal discovery, formulating domain hypotheses and identifying both direct and causal proxy signals validated against published literature. PPE systematically retrieves covariates from established geospatial repositories including Data Commons and Google Earth Engine . For signals not available in established repositories, the system performs live open-web discovery, searching government portals and academic repositories at inference time. Stage 2: Multimodal dataset curation. The assembled 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. The system implements a strict “Feature Gate” that enforces automated target leakage mitigation, evaluating every candidate covariate against four anti-leakage criteria (filtering out mathematical sub-components, shared survey data, downstream causal effects, and future temporal data) to ensure the integrity of downstream evaluations. Stage 3: Automated model building and prediction. Here, PPE shifts from data curation to model optimization and evaluation. The system searches over multiple model families such as regularized linear models, gradient-boosted decision trees (GBDT), and multi-layer perceptrons , with custom logic to prevent overfitting (Overfit