Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings

Evelyn Ma; Rama Kumar Pasumarthi; Kishwar Shafin; Mandar Sharma; Mimi Sun; Hamed Sadeghi; Dav M. Ebengo; Mbulayi Onesime

cs.AIAug 26, 2026

Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings

Authors: Evelyn Ma, Rama Kumar Pasumarthi, Kishwar Shafin, Mandar Sharma, Mimi Sun, Hamed Sadeghi, Dav M. Ebengo, Mbulayi Onesime,

Organizations: Google Research · Institut National de Recherche Biomédicale, Democratic Republic of Congo · University of Oxford

Abstract

Addressing critical global challenges, from food security and disaster risk to disease outbreaks and socio-economic vulnerability, demands high-fidelity geospatial modeling. However, building predictive planetary models remains bottlenecked by a fragmented data ecosystem, requiring manual data retrieval, multimodal data curation and fusion along with iterative model selection. We present the Planetary Prediction Engine (PPE), an autonomous AI system that executes this end-to-end workflow directly from natural-language queries. PPE synthesizes multimodal datasets on the fly, retrieving spatiotemporally relevant covariates across open-web and Earth observation platforms (Data Commons, Google Earth Engine) and fusing them with geospatial foundation model embeddings (PDFM, AlphaEarth). Simultaneously, it searches over task-tailored model architecture families with automated overfitting guards. Across diverse tasks, geographies, and scientific domains, PPE consistently outperforms state-of-the-art or manually tuned expert baselines. For US spatial regression, PPE improves mean $R^2$ across 21 CDC health indicators (76.8% vs. 60.0%), FEMA national risk indices (64.9% vs. 60.0%), and the Social Vulnerability Index (66.2% vs. 58.6%). For spatial downscaling in data-scarce settings, PPE integrates localized proxies to double baseline accuracy in Nigerian food security indicators ( $R^2$ of 66.1% vs. 31.5%). For epidemiological nowcasting of the 2026 DRC Bundibugyo Ebola outbreak, PPE achieves a Recall@10 of 83.3% (identifying 15 of 18 newly invaded health zones across five weekly forecasts), a +10.3 percentage-point improvement over the public state-of-the-art modeling (~73%). By combining autonomous multimodal planetary data discovery with targeted model optimization, PPE lowers the technical barrier to planetary-scale analytics, enabling rapid, customized, expert-level deployment.

Figures & tables

Figure 1: The Planetary Prediction Engine ’s end-to-end workflow. The system decomposes the predictive workflow into three modular stages: (1) Intelligent data selection, (2) dataset curation, and (3) AutoML & prediction to produce the predictions and report. Off-the-shelf LLMs serve as orchestrators within each stage.

ML Paradigm	Geographic Scope	Thematic Domain	Benchmark Dataset	Prediction Targets	Train / Test Granularity	Train / Test Size
Nowcasting	Global South (DRC)	Humanitarian Epidemics	DRC Ebola Outbreak Tracking	one-week confirmed caseloads increase	Admin 3 health zones	7 weeks $\times$ 519 zones / 5 weeks $\times$ 519 zones
Super-Resolution Downscaling	Global South (Nigeria)	Food Security	Nigeria FCG Insecurity	FCG score	Admin 1 $\to$ Admin 2 (LGA)	30 states $\times$ 40 months / 581 LGAs $\times$ 40 months
Super-Resolution Downscaling	Global North (US)	Social Vulnerability	SVI	5 index scores	County $\to$ ZCTA	$\sim$ 3k / $\sim$ 33k
Spatial Regression	Global North (US)	Public Health	CDC Health	21 health variables	Census Tract	$\sim$ 67k / $\sim$ 17k
	Global North (US)	Environmental Risk	FEMA NRI	21 risk scores	Census Tract	$\sim$ 67k / $\sim$ 17k
	Global North (US)	Social Vulnerability	SVI	5 index scores	County	$\sim$ 2.4k / $\sim$ 0.6k

Table 1: Summary of Evaluation Benchmarks and Dataset Specifications. Overview of target variables, geographic scopes, and spatial resolutions across the three evaluation paradigms.

System	Feature Configuration	Data Preprocessing (DP)	Recall@10 (%)[95% CI]
Baseline/ SOTA (Bayesian modeling)	Baseline Signals	Baseline DP	$\sim 73$
PPE (Covariates)	Baseline Signals + Geospatial Covariates	Baseline DP + PPE automatic tabular DP	77.8 [54.8, 91.0]
PPE (Full Stack)	Baseline signals + PPE automatic full data selection	Baseline DP + PPE automatic full DP	83.3 [60.8, 94.2]

Table 2: Model Performance (Recall@10) across Feature Configurations on tracking the spatial invasion of Bundibugyo ebolavirus across the Democratic Republic of the Congo (DRC).

Panel of Figure 2 — Figure 2: Frontline Ebola Transmission prediction visualization: Comparison of predicted total caseloads against ground-truth caseloads.

Figure 3: Super-Resolution Food Security Downscaling in Nigeria (ADM1 State Level $\to$ ADM2 LGA Level)

System	Feature Configuration	Overall $R^{2}$ [95%CI] (%)
Baseline	Macro-Covariates + Interpolation	31.5 [19.0, 39.3]
PPE (Covariates)	Macro-Covariates + nighttime lights (NTL)	49.8 [39.8, 57.9]
PPE (Covariates)	Macro-Covariates + vegetation	60.1 [40.3, 72.8]
PPE (Full Stack)	Macro-Covariates + vegetation + NTL + Automatic Intelligent Selection	66.1 [55.9, 72.8]

Table 3: Model Performance (

R^{2}

) for Super-Resolution Food Security Downscaling (FCG) in Nigeria (ADM1 State Level

\to

ADM2 LGA Level).

System	Feature Configuration	Mean $R^{2}$ [95% CI]
Baseline	Macro-Covariates + Interpolation	11.0 [10.0, 12.3]
PPE (Covariates)	Geospatial Covariates	25.6 [24.3, 26.9]
PPE (Embeddings)	PDFM	36.9 [36.0,37.9]
PPE (Full Stack)	Covariates + PDFM	37.6 [36.7, 38.6]

Table 4: Model Performance (

R^{2}

) across Feature Configurations on SVI (Social Vulnerability Index) Benchmark Themes with Super-resolution tasks (from county-level to zipcode-level).

System	Feature Configuration	Mean $R^{2}$ [95% CI] (%)
Baseline / SOTA	Manual Expert Pipeline (PDFM + AEF)	60
PPE (Embeddings)	PDFM	59.7 [58.6, 60.9]
PPE (Embeddings)	PDFM+AEF	61.8 [60.6, 62.9]
PPE (Full Stack)	PDFM + AEF + Data Commons Covariates	76.8 [76.1, 77.6]

Table 5: Model Performance (

R^{2}

) across Feature Configurations on 21 CDC Health Variables.

System	1. Socioeconomic & Composite	2. Atmospheric & Climatological	3. Geophysical & Hydrological	Full Nationwide Suite
Target Count	4	10	6	20
PDFM + AEF Manual expert Mean $R^{2}$ (%)	61.1	64.6	51.1	59.9
PDFM + AEF + DC (Feature ablation) Mean $R^{2}$ $[95\%\text{ CI}]$ (%)	66.9 [66.2, 67.5]	64.3 [63.6, 65.0]	51.7 [49.3, 53.7]	61.1 [60.2, 61.7]
PDFM + AEF + DC + Intelligent Data Selection Mean $R^{2}$ $[95\%\text{ CI}]$ (%)	69.4 [68.8, 70.0]	68.3 [67.5, 68.9]	56.2 [54.1, 57.8]	64.9 [64.1, 65.5]

Table 6: Model Performance (

R^{2}

) across Feature Configurations on FEMA Environmental Risk Variables.

System	Feature Configuration	Mean $R^{2}$ [95% CI] (%)
Baseline	Foundation Model Signals	60.3 [55.2, 64.9]
PPE (Covariates)	Geospatial Covariates	50.2 [44.2, 55.6]
PPE (Full Stack)	Covariates + PDFM	66.2 [61.6, 70.4]

Table 7: Model Performance (

R^{2}

) across Feature Configurations on SVI Benchmark Themes (Spatial Regression at County-Level).

Panel of Figure 4 — Figure 4: The Planetary Prediction Engine’s multi-stage modular architecture.

Model	Class	Implementation	Key Hyperparams
linear	Linear	Scikit-Learn (Ridge, Lasso, ElasticNet) [ 57 ]	Regularization Type, Penalty $\alpha$ , L1 ratio $\rho$
boost	HistBoost, XGBoost [ 12 ]	Scikit-Learn HistGradientBoosting Regressor; xgboost.XGBRegressor	Learning rate, Max leaf nodes,Min samples per leaf, Loss Type,Max tree depth, # of estimators,Subsample, Regularization
mlp	MLPKeras	Keras	Hidden layer sizes, Dropout rate, Optimizer configs

Table 8: Model types supported by the Planetary Prediction Engine.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Covariate Category	Data Source / Provider	Key Variables & Dimensions Captured	Access Class
Socio-Demographic & Economic	Data Commons (Google Open Source)	Population density, household income trends, poverty rates, age/gender distributions, chronic disease rates.	Public (Open API & Free)
Ecological & Environmental	Google Earth Engine (EE)	Multi-spectral imagery (Sentinel-2, Landsat, MODIS), NDVI (vegetation), soil moisture, LST anomalies.	Public (Free for Research Use)
Infrastructure & Mobility	Google Maps Platform Insights	Aggregated Statistics on POIs	Public (GCP Commercial Billing)
Infrastructure & Mobility	OpenStreetMap (OSM) (via OCHA HDX)	Road network vectors, administrative boundary polygons, structural building footprints.	Public (Open Source / Free)
Meteorological	Open-Meteo / NOAA	Daily temperature, cumulative precipitation, relative humidity, wind velocity profiles.	Public (Open API / Free)

Appendix

Table 9: Inventory of Geospatial Covariates

Foundation Model	Population Dynamics Foundation Model (PDFM)	AlphaEarth Foundation Model
Modality / Coverage	Socio-economic & Demographic (State, County, ZCTA, City)	Ecological & Remote Sensing (High-resolution satellite raster)
Dimension	330 / 512	64
Key Semantic Characteristics	Encodes latent population density, mobility, economic growth, and regional demographics.	Encodes multi-scale topography, vegetation canopy, land-use semantics, and terrain profiles.
Access Class	Pre-General Availability + Research Use	Publicly available

Appendix

Table 10: Inventory of Geospatial Foundation Models

Dimension	Weight	Criteria & Point Allocations
Provenance & License	High (x5)	Institutional backing and license openness. Openly licensed data from established repositories (4 pts) → government/NGO open data (3 pts) → attribution-required academic data (2 pts) → restrictive or unknown license (0 pts).
Spatio-Temporal Fitness	Medium (x2)	Spatial coverage of the target area of interest and temporal overlap with the prediction window.
Signal Alignment	High (x3)	Degree to which the dataset directly addresses a signal from the Signal Guide: exact variable match (4 pts) → requires transformation (3 pts) → proxy coverage (2 pts) → tangential (0 pts).
Format & Quality	Low (x1)	Data readiness: structured and pre-parsed (4 pts) → standard interoperable format (2 pts) → legacy or malformed (0 pts).
Redundancy	Low (x1)	Whether the dataset provides unique coverage or duplicates a higher-scoring source for the same signal.

Appendix

Table 11: Five-Dimension Scoring Rubric

FEMA Target Variable	PDFM + AEF Manual expert Mean R²	PDFM + AEF + DC (Feature ablation) Mean R² [95% CI]	Data Common variables used (Comma-Separated Variables)
Social Vulnerability	0.4824	0.6755 [0.6556, 0.6941]	Count_Person, BelowPovertyLevelInThePast12Months, Count_Household, HouseholderAge65OrMoreYears, SingleMotherFamilyHousehold, LimitedEnglishSpeakingHousehold, NoComputer, NoInternetAccess, With0AvailableVehicles, WithFoodStampsInThePast12Months, IncomeOfUpto10000USDollar, IncomeOf10000To14999USDollar, IncomeOf15000To19999USDollar, IncomeOf20000To24999USDollar, IncomeOf100000To124999USDollar, IncomeOf200000OrMoreUSDollar, Count_HousingUnit, Count_HousingUnit_Before1939DateBuilt
Resilience Score	0.7402	0.7358 [0.7287, 0.7427]	Count_Person, BelowPovertyLevelInThePast12Months, Count_Household, HouseholderAge65OrMoreYears, SingleMotherFamilyHousehold, LimitedEnglishSpeakingHousehold, NoComputer, NoInternetAccess, With0AvailableVehicles, WithFoodStampsInThePast12Months, IncomeOfUpto10000USDollar, IncomeOf10000To14999USDollar, IncomeOf15000To19999USDollar, IncomeOf20000To24999USDollar, IncomeOf100000To124999USDollar, IncomeOf200000OrMoreUSDollar, Count_HousingUnit, Count_HousingUnit_Before1939DateBuilt
Composite Risk Score	0.5983	0.6221 [0.6078, 0.6335]	Count_Person, BelowPovertyLevelInThePast12Months, Count_Household, HouseholderAge65OrMoreYears, SingleMotherFamilyHousehold, LimitedEnglishSpeakingHousehold, NoComputer, NoInternetAccess, With0AvailableVehicles, WithFoodStampsInThePast12Months, IncomeOfUpto10000USDollar, IncomeOf10000To14999USDollar, IncomeOf15000To19999USDollar, IncomeOf20000To24999USDollar, IncomeOf100000To124999USDollar, IncomeOf200000OrMoreUSDollar, Count_HousingUnit, Count_HousingUnit_Before1939DateBuilt
Expected Annual Loss	0.6227	0.6423 [0.6319, 0.6531]	Count_Person, BelowPovertyLevelInThePast12Months, Count_Household, HouseholderAge65OrMoreYears, SingleMotherFamilyHousehold, LimitedEnglishSpeakingHousehold, NoComputer, NoInternetAccess, With0AvailableVehicles, WithFoodStampsInThePast12Months, IncomeOfUpto10000USDollar, IncomeOf10000To14999USDollar, IncomeOf15000To19999USDollar, IncomeOf20000To24999USDollar, IncomeOf100000To124999USDollar, IncomeOf200000OrMoreUSDollar, Count_HousingUnit, Count_HousingUnit_Before1939DateBuilt

Appendix

Table 12: PPE performance of FEMA targets (Socioeconomic & Composite) with ablation on PDFM + AEF + DC features

FEMA Target Variable	PDFM + AEF Manual expert Mean R²	PDFM + AEF + DC (Feature ablation) Mean R² [95% CI]	Data Common variables used (Comma-Separated Variables)
Tornado Risk	0.7543	0.7542 [0.7468, 0.7617]	None
Hail Risk	0.5949	0.5976 [0.5807, 0.6138]	None
Strong Wind Risk	0.7339	0.7271 [0.7179, 0.7362]	None
Lightning Risk	0.4884	0.4918 [0.4758, 0.5090]	None
Drought Risk	0.6700	0.6866 [0.6332, 0.7323]	None
Heat Wave Risk	0.6336	0.6350 [0.6193, 0.6489]	None

Appendix

Table 13: PPE performance of FEMA targets with ablation on PDFM + AEF + DC features

FEMA Target Variable	PDFM + AEF Manual expert Mean R²	PDFM + AEF + DC (Feature ablation) Mean R² [95% CI]	Data Common variables used (Comma-Separated Variables)
Coastal Flooding	0.3786	0.3840 [0.3394, 0.4240]	None
Riverine Flooding	0.2930	0.2935 [0.2695, 0.3179]	None
Hurricane Risk	0.8270	0.8124 [0.8011, 0.8232]	None
Earthquake Risk	0.8606	0.8608 [0.8540, 0.8672]	None
Landslide Risk	0.5614	0.5624 [0.5339, 0.5887]	None
Avalanche Risk	0.1474	0.1866 [0.0572, 0.2947]	Count_Person, BelowPovertyLevelInThePast12Months, Count_Household, HouseholderAge65OrMoreYears, SingleMotherFamilyHousehold, LimitedEnglishSpeakingHousehold, NoComputer, NoInternetAccess, With0AvailableVehicles, WithFoodStampsInThePast12Months, IncomeOfUpto10000USDollar, IncomeOf10000To14999USDollar, IncomeOf15000To19999USDollar, IncomeOf20000To24999USDollar, IncomeOf100000To124999USDollar, IncomeOf200000OrMoreUSDollar, Count_HousingUnit, Count_HousingUnit_Before1939DateBuilt

Appendix

Table 14: PPE performance of FEMA targets (Geophysical & Hydrological) with ablation on PDFM + AEF + DC features

FEMA Target Variable	PDFM + AEF + DC Feature Ablation Mean R²	PDFM + AEF + Discovered DC Variables Intelligent Feature Selection Mean R² [95% CI]	Discovered / Selected External Variables (Sourced Non-Linearly)
Social Vulnerability	0.6755	0.6773 [0.6698, 0.6869]	Count_Person, HouseholderAge65OrMoreYears, SingleMotherFamilyHousehold
Resilience Score	0.7358	0.7258 [0.7062, 0.7348]	LimitedEnglishSpeakingHousehold, Median household income
Composite Risk Score	0.6221	0.6749 [0.6641, 0.6861]	WithFoodStampsInThePast12Months, Health insurance rates
Expected Annual Loss	0.6423	0.6974 [0.6888, 0.7055]	Count_HousingUnit_Before1939DateBuilt, Poverty Index metrics

Appendix

Table 15: PPE performance on FEMA Socioeconomic & Composite risk targets with ablation on PDFM + AEF + DC and intelligent data selection.

FEMA Target Variable	PDFM + AEF + DC Feature Ablation Mean R²	PDFM + AEF + Discovered DC Variables Intelligent Feature Selection Mean R² [95% CI]	Discovered / Selected External Variables (Sourced Non-Linearly)
Tornado Risk	0.7542	0.9168 [0.9123, 0.9208]	Agricultural vs Developed land ratios, Surface terrain elevation, Annual average precipitation, Distance to coast lines, Relative humidity
Hail Risk	0.5976	0.8656 [0.8540, 0.8807]	Cloud Cover fractions, Vegetation indices (NDVI) variance, Land surface temperature gradients, Distance to major water bodies
Strong Wind Risk	0.7339	0.6891 [0.6631, 0.7117]	NOAA Convective Storm Frequencies, Spherical coordinates (lat/lon 3D embeddings), Atmospheric pressure gradient proxies
Lightning Risk	0.4884	0.6607 [0.6051, 0.7055]	US Drought Monitor Palmer Indices, Agricultural land fraction, Water body area ratio, Mean annual precipitation
Drought Risk	0.6700	0.8711 [0.8648, 0.8775]	NOAA High Wind Event frequencies, Coastal proximity, Terrain elevation, Topographic wind exposure
Heat Wave Risk	0.6336	0.5964 [0.5790, 0.6139]	US Census Educational attainment, Health insurance coverage fraction, Median household income, Internet access

Appendix

Table 16: PPE performance on FEMA Atmospheric & Climatological risk targets with ablation on PDFM + AEF + DC and intelligent data selection.

FEMA Target Variable	PDFM + AEF + DC Feature Ablation Mean R²	PDFM + AEF + Discovered DC Variables Intelligent Feature Selection Mean R² [95% CI]	Discovered / Selected External Variables (Sourced Non-Linearly)
Coastal Flooding	0.3786	0.5047 [0.4882, 0.5202]	USGS Fault Line distance proxies + AlphaEarth seismic geodynamics embeddings
Riverine Flooding	0.2930	0.3144 [0.2924, 0.3377]	Coastal proximity buffer, Historical tropical cyclone track density, Elevation
Hurricane Risk	0.8270	0.7729 [0.7650, 0.7809]	AlphaEarth satellite foundation embeddings + discovered vegetative fuel covariates
Earthquake Risk	0.8606	0.7992 [0.7923, 0.8059]	Mean minimum temperature, Sub-zero degree day counts, Latitude spatial embeddings
Landslide Risk	0.5614	0.7547 [0.7385, 0.7705]	US Census Gazetteer spatial geometry, Census Tract Housing Age, Income distributions, Building Density, Population Density
Avalanche Risk	0.1474	0.6882 [0.6767, 0.6988]	NOAA Severe Hail probability metrics, Convective available potential energy (CAPE) proxies, Terrain elevation

Appendix

Table 17: PPE performance on FEMA Geophysical & Hydrological risk targets with ablation on PDFM + AEF + DC and intelligent data selection.

Automated Machine Learning

Automated machine learning (AutoML) builds and tunes data-to-model pipelines with minimal manual effort, balancing predictive quality, compute, and deployment constraints.

See all 97 papers

Oct 5, 2026

Planetary Geospatial Foundation Models: A New Paradigm for Global Public Health

The efficacy of traditional disease prediction is limited by spatial gaps and temporal lags, which impact the timing and targets of resource deployments. Outbreaks escalate undetected, chronic disease burdens are quantified years later, and at-risk populations in data-sparse regions remain unaddressed. Planetary geospatial foundation models complement existing epidemiological workflows to provide operational improvements, encoding multimodal search, mobility, and environmental signals into generalizable place representations. As illustrations of this complementarity, we present independent global health case studies of Google Earth AI's Population Dynamics Foundation Model (PDFM) -- a foundation model for geospatial inference -- across four domains (vaccine-preventable, communicable, noncommunicable, maternal mental health), five tasks (spatial extrapolation, interpolation/nowcasting, probabilistic forecasting, prospective forecasting, risk stratification), and four countries (USA, Canada, Mexico, and the Democratic Republic of the Congo). Across these case studies, PDFM addresses critical surveillance gaps across domains: improving US-Canada border MMR vaccination coverage predictions by capturing cross-border behavioral spillovers domestic models miss; nowcasting cardiovascular disease to accelerate data availability; enhancing short-term municipal Mexican dengue forecasts for timely outbreak vector control; improving forecasts of cholera hotspots; and adding a transferable signal to individual-level postpartum-depression risk prediction in US states the model had never seen, while not replacing individual socioeconomic data or closing demographic screening gaps. Together, these results showcase capabilities of geospatial foundation models for public health surveillance.

Arbaaz Muslim, Aviv Slobodkin, Katherine Wheeler-Martin +35

Google, Mountain View, CA, USA · Department of Population Health, NYU Grossman School of Medicine, New York, NY, USA · Center for Child Health Services Research, Icahn School of Medicine at Mount Sinai, New York, NY, USA +14

Epidemic Forecasting Geospatial Foundation Models Spatiotemporal Forecasting

Ask PDF

Mar 7, 2026

Self-Supervised Multi-Modal World Model with 4D Space-Time Embedding

We present DeepEarth, a self-supervised multi-modal world model with Earth4D, a novel planetary-scale 4D space-time positional encoder. Earth4D extends 3D multi-resolution hash encoding to include time, efficiently scaling across the planet over centuries with sub-meter, sub-second precision. Multi-modal encoders (e.g. vision-language models) are fused with Earth4D embeddings and trained via masked reconstruction. We demonstrate Earth4D's expressive power by achieving state-of-the-art performance on an ecological forecasting benchmark. Earth4D with learnable hash probing surpasses a multi-modal foundation model pre-trained on substantially more data. Access open source code and download models at: https://github.com/legel/deepearth

Lance Legel, Qin Huang, Brandon Voelker +9

Ecological Intelligence Lab Ecodash.ai · School of Complex Adaptive Systems Arizona State University · Geosensing Systems Engineering & Sciences Lab University of Houston +6

World Model Learning Geospatial Representation Learning Positional Encoding Multimodal Pretraining

Ask PDF

Sep 17, 2026

JEPA-Anything: Learning Predictive Models across Different Worlds

World modeling enables intelligence to anticipate consequences, guide interventions, and learn from interaction. Yet predictive models remain domain-specific: can a common learning principle support world modeling across radically different systems? We introduce JEPA-Anything, a domain-agnostic framework based on orthogonal predictive factorization (OPF). Extending joint-embedding predictive architectures, OPF decomposes latent targets into complementary factors, learns them through dedicated pathways, and recombines them within a shared predictive design. We evaluate JEPA-Anything across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather. Experiments span representation learning, intervention prediction, out-of-distribution generalization, and long-horizon dynamics, including 10 matched dynamics tasks, forecasting of over 1,000 clinical events, and 100-step molecular rollouts across four systems. Against matched JEPA baselines, JEPA-Anything improves reported metrics on all 10 dynamics tasks and reduces single-intervention prediction error on Interventional Pong by 34.8%. It achieves the lowest one-step and 100-step molecular errors among compared methods in all four systems. Beyond prediction, a factor-nominated biological intervention receives experimental support in cell co-cultures, patient-derived organoids, tumor fragments, and mice; latent orbital modes recover the Keplerian scaling exponent with a fitted slope of -1.4991. These results support a common factorized predictive principle across heterogeneous worlds, connecting world modeling with intervention and experimentally grounded scientific discovery. Code: https://github.com/Gen-Verse/JEPA-Anything

Taoyong Cui, Zhongyao Wang, Xinyue Xu +10

World Model Learning Joint-Embedding Predictive Architecture Predictive Representation Learning

Ask PDF

Abstract

Automated Machine Learning

Explore similar work

Planetary Geospatial Foundation Models: A New Paradigm for Global Public Health

Self-Supervised Multi-Modal World Model with 4D Space-Time Embedding

JEPA-Anything: Learning Predictive Models across Different Worlds