Population Dynamics Foundation Model (PDFM)

• The PDFM framework is a versatile and reusable representation-learning and population dynamics simulation for forecasting and inferring populations in spatial, demographic, ecological, and neural domains
• It encompasses a variety of foundation models that encode diverse population features, such as age distributions, phenotype distributions, geospatial contexts and mechanistic mechanistic constraints, facilitating predictions on downstream health, socioeconomic, environmental, and forecasting applications but is note a demographic simulator per se.
• Popular applications of PDFM include health indicators, socioeconomic factors, geographical interpolation, extrapolation, forecasting tasks, and actionable insights

The Population Dynamics Foundation Model (PDFM) is a class of reusable representation-learning and mechanistic-modeling systems for forecasting, inferring, and simulating populations across spatial, temporal, demographic, biological, ecological, and neural domains. The term is used most concretely for a geospatial foundation model that encodes aggregated human behavior, built-environment structure, mobility, weather, and air quality into location embeddings for downstream health, socioeconomic, environmental, and forecasting tasks (Agarwal et al., 2024). More broadly, a PDFM can denote a hybrid architecture that represents population states as densities, age distributions, phenotype distributions, or probability measures; learns transition operators from heterogeneous observations; preserves mechanistic constraints; and supports transfer across populations, environments, and intervention settings.

1. Conceptual scope and terminology

PDFM does not denote a single universally standardized architecture. The most operationally established usage is the Population Dynamics Foundation Model developed for general geospatial inference. It produces reusable geographic embeddings through a graph neural network (GNN), using postal codes and counties as geographic units and combining aggregated search trends, maps and points of interest, busyness, weather, and air quality (Agarwal et al., 2024). Its downstream applications include health indicators, socioeconomic factors, environmental measurements, interpolation, extrapolation, super-resolution, and forecasting.

A broader technical interpretation treats PDFM as a population-level analogue of a foundation model: a system pretrained or calibrated across heterogeneous populations and contexts, then adapted to forecasting, classification, simulation, inverse inference, or intervention optimization. Under this interpretation, population state may be represented as:

• a spatial density field, such as $u(x,t)$ or $\rho(\mathbf{x},t)$;
• an age distribution, such as $p(t,a)$;
• a biological-age distribution, such as $u(t,b)$ or $u(t,a,b)$;
• a phenotype-structured density, such as $m(\phi,t)$ or $m(\phi,r,t)$;
• a probability measure over agent states;
• a discrete population-count vector or reaction-network state;
• a neural population-token matrix.

This broader formulation is supported by several complementary research directions. A first-principles framework represents populations through reaction networks, continuous-time Markov jump processes, deterministic density limits, and Bayesian inference (Araujo, 2023). A controlled mean-field model represents chiplet populations as probability densities evolving under nonlocal stochastic interactions and electrode controls (Nodozi et al., 2023). A calcium-imaging model, CalM, uses self-supervised tokenization and dual-axis autoregressive attention to model neural population activity (Xu et al., 3 Apr 2026). A latent Fokker–Planck approach learns distributional dynamics from snapshots without individual trajectories (Huang et al., 24 Jul 2026).

The defining distinction is therefore between population representation and population dynamics. A geospatial PDFM primarily learns a reusable representation of place, whereas a full population-dynamics foundation model would additionally model temporal evolution, uncertainty, interventions, conservation laws, latent states, and population-level transition mechanisms.

2. Geospatial PDFM architecture and representation learning

The geospatial PDFM constructs representations for geographic units and their graph context. In the principal United States implementation, nodes comprise postal codes and counties. Bidirectional proximity edges connect postal-code or county centroids within 100 miles, while containment edges connect postal codes and counties whose boundaries overlap. The graph can also encode similarity in behavioral features such as aggregated search trends (Agarwal et al., 2024).

The input modalities are:

1. Aggregated search trends: normalized frequencies for the 1,000 most common queries, with query texts removed from subsequent analysis.
2. Maps and points of interest: counts across 1,192 common Google Maps categories.
3. Busyness: aggregated visits to locations associated with map categories.
4. Weather: summaries of meteorological variables.
5. Air quality: summaries including carbon monoxide, nitrogen dioxide, ozone, sulfur dioxide, particulate matter, and air-quality index.

The model applies column-wise standardization, clipping of extreme values, modality concatenation, and projection through a hidden layer with GeLU activation. Training uses GraphSAGE-style subgraph sampling. Each seed node is associated with sampled postal-code and county neighbors up to four hops away, while the implemented GNN uses one message-passing round. A linear layer produces a 330-dimensional embedding.

The embedding is partitioned by source:

• dimensions 0–127 encode aggregated search trends;
• dimensions 128–255 encode maps and busyness;
• dimensions 256–329 encode weather and air quality.

Foundation-model pretraining is self-supervised. Each embedding partition reconstructs its corresponding input modality using a Huber loss. This objective encourages the representation to retain information from heterogeneous sources without requiring downstream health, economic, or environmental labels. The fixed embeddings are subsequently used with ridge regression, multilayer perceptrons, gradient-boosted decision trees, or forecasting models.

The representation is spatially contextual rather than purely local. It combines a node’s own observations with neighboring and related locations, thereby encoding geographic proximity, administrative containment, and behavioral similarity. This structure supports transfer to locations with sparse labels, although the embedding remains tied to the geographic support and graph construction used during representation learning.

A later global formulation maps Search Trends to Google Knowledge Graph entities rather than retaining raw query counts. It covers 17 countries and uses postal codes or similarly sized regions, counties, administrative regions, census tracts, and NUTS 3 regions. Monthly embeddings extend the static representation toward temporal modeling, although the reported temporal coverage is approximately two years and the temporal embeddings use a different Search Trends preprocessing procedure from the global static embeddings (Bell et al., 21 Oct 2025).

The resulting representation is privacy-aware but not formally differentially private. Aggregation, frequency thresholds, anonymized busyness data, regional embeddings, and removal of query text reduce direct identification risk. Formal differential-privacy guarantees, quantified re-identification bounds, and comprehensive privacy audits are not established in the cited work.

3. Population-state representations and mechanistic foundations

A full PDFM requires state representations that preserve the structure relevant to the population under study. Age-structured demography provides one such representation. The PAT model uses a discrete population-age-time state $p(t,a)$, with age classes $a=0,1,\ldots,A$. Individuals advance one age class per time step subject to age-dependent mortality, while newborns enter the age-zero class according to fertility contributions from reproductive ages (Feng et al., 2021). The total population is:

$P(t)=\sum_{a=0}^{A}p(t,a).$

The model’s population balance is births minus deaths. Its stationary age profile satisfies a replacement condition in which age-specific reproductive contributions are discounted by survival to reproductive ages. Numerical experiments identify subreplacement, replacement, and superreplacement regimes, with an approximately two-children-per-woman threshold in the model’s parameterization. The replacement regime produces a kink-shaped age distribution generated by relatively high survival at young ages and increasing mortality at older ages.

Biological-age modeling extends this representation by distinguishing chronological age $\rho(\mathbf{x},t)$0 from biological age $\rho(\mathbf{x},t)$1. Individuals age continuously in $\rho(\mathbf{x},t)$2 but undergo stochastic rejuvenating or premature-aging events. The density $\rho(\mathbf{x},t)$3 obeys a transport-and-jump equation with mass conservation in the base model (Demongeot et al., 2023). Adding chronological age produces a joint density $\rho(\mathbf{x},t)$4, allowing the model to represent age-dependent event rates, mortality, cohort effects, and birth distributions. This structure is relevant to PDFM design because biological age is a latent state inferred from biomarkers, clinical events, treatment history, and environmental exposures rather than a directly observed scalar.

Phenotype-structured tumor models provide another state representation. In the melanoma MITF rheostat model, the continuous phenotype coordinate $\rho(\mathbf{x},t)$5 distinguishes invasive, proliferative, and differentiated states. A density $\rho(\mathbf{x},t)$6 describes the abundance of cells at each phenotype, while $\rho(\mathbf{x},t)$7 additionally resolves radial position (Chambers et al., 13 Jul 2026). The population model couples phenotype drift and diffusion to phenotype-dependent proliferation, death, density-dependent stress, and spatial motility. This structure demonstrates that reversible individual transitions can produce irreversible population-level distributions through selection, feedback, hysteresis, and multiple attractors.

Reaction-network formulations provide a discrete mechanistic ontology. A population state is a count vector $\rho(\mathbf{x},t)$8, and reactions transform it through stoichiometric displacement vectors $\rho(\mathbf{x},t)$9. The master equation describes stochastic transitions, while the deterministic density limit yields:

$p(t,a)$0

This hierarchy connects individual events, stochastic population trajectories, deterministic equations, fluctuations, and Bayesian parameter inference (Araujo, 2023). It also exposes an identifiability limitation: multiple reaction networks can generate the same macroscopic differential equation.

Probability-density representations support populations whose individual identities are unavailable or unnecessary. In the chiplet mean-field model, the population is a normalized density $p(t,a)$1 in $p(t,a)$2, with finite second moments. The density satisfies a controlled, nonlinear, nonlocal Fokker–Planck equation and evolves as a Wasserstein gradient flow of an interaction-plus-entropy functional (Nodozi et al., 2023). In snapshot-only learning, a latent distribution $p(t,a)$3 evolves through a Fokker–Planck model and is transformed into the observed distribution $p(t,a)$4 by an invertible transport map:

$p(t,a)$5

This latent–transport decomposition permits distribution forecasting without temporal correspondence between individuals (Huang et al., 24 Jul 2026).

4. Temporal, spatial, and stochastic dynamics

The dynamics of a PDFM must distinguish biological response times, diffusion or transport times, switching times, and observation times. In an intermittent-refuge reaction–diffusion model, the population density $p(t,a)$6 obeys a Fisher–KPP/logistic equation with spatially localized and temporally blinking protection (Colombo et al., 2016):

$p(t,a)$7

The refuge has length $p(t,a)$8, the harmful-effect rate is $p(t,a)$9, and the duty-cycle parameter $u(t,b)$0 specifies the fraction of each period during which the harmful effect penetrates the refuge. The total population is $u(t,b)$1, and the sign of the cycle-averaged per-capita growth rate determines extinction or persistence. In the fast-switching regime, $u(t,b)$2, the refuge experiences an effective internal growth rate $u(t,b)$3. In the slow-switching regime, population growth responds separately to harmful and protected intervals. Spatial memory, boundary flux, and an external population reservoir make slow switching more favorable than fast switching at fixed harmful duty cycle.

This model illustrates why temporal averages are insufficient for population forecasting. The outcome depends on refuge geometry, diffusion, switching period, spatial gradients, and the history of redistribution. A PDFM intended for intervention optimization must therefore encode both current population state and relevant historical context.

Spatially resolved population systems can have qualitatively different dynamics from well-mixed models. In the melanoma model, high-motility invasive cells occupy dense, stressed regions behind a proliferative or differentiated wavefront, while the leading edge is typically dominated by proliferative and differentiated phenotypes. Spatial coupling can localize oscillations near the tumor core while producing an approximately steady outward wave at larger radii (Chambers et al., 13 Jul 2026). Similar distinctions occur in ecological refuges, geospatial populations, and controlled mean-field systems.

Stochasticity enters at multiple levels. Reaction-network models distinguish demographic event noise from deterministic density limits. Mean-field chiplet models distinguish individual Brownian fluctuations from the deterministic density evolution. Biological-age models represent event-time and event-type randomness through piecewise-deterministic Markov processes. CalM represents neural activity through a discrete neural-trace vocabulary and autoregressive stochastic forecasting. Snapshot-based Fokker–Planck learning represents stochasticity through latent diffusion and distributional prediction.

A PDFM should therefore distinguish:

• process noise: stochasticity intrinsic to population transitions;
• observation noise: measurement uncertainty and sampling error;
• parameter uncertainty: uncertainty in rates, kernels, and mechanisms;
• representation uncertainty: uncertainty caused by incomplete or biased modalities;
• finite-population fluctuations: deviations from continuum or mean-field limits.

5. Learning objectives, transfer, and downstream tasks

The geospatial PDFM uses fixed embeddings as reusable covariates. Its benchmark includes 27 spatial tasks covering 21 health indicators, four socioeconomic variables, and two environmental variables. It achieves a reported mean $u(t,b)$4 of 0.83 across interpolation tasks, 0.70 across extrapolation tasks, and a mean within-county Pearson correlation of 0.48 for super-resolution. The reported model is best on all interpolation tasks and on 25 of 27 extrapolation and super-resolution tasks, although IDW remains competitive for smoothly varying variables such as elevation and satellite-based models perform better for some environmental targets (Agarwal et al., 2024).

The model also improves TimesFM forecasts when a location embedding is supplied through a lightweight adapter. For county unemployment, TimesFM with PDFM embeddings achieves MAPE 0.1132 compared with 0.1186 for TimesFM alone. For postal-code poverty, the corresponding values are 0.2577 and 0.3215. These results indicate that static spatial context can correct systematic errors in univariate temporal forecasts.

External studies show that performance is conditional. In Malawi, PDFM embeddings were compared with AlphaEarth, mobile-phone call-detail-record embeddings, and spatial interpolation for 15 health-program indicators across 552 catchments. Embedding-based methods improved on geostatistical baselines for 13 of 15 indicators, while a Multi-GeoFM model achieved test-set $u(t,b)$5 values of 0.64 for population density, 0.68 for new HIV cases, and 0.55 for child vaccinations (Metz et al., 29 Oct 2025). TB and malnutrition remained difficult, reflecting sparse and inconsistent outcome data.

A population-estimation study in Brazil, Nigeria, and the United States found that fixed 330-dimensional PDFM embeddings reduced unexplained variance by a median 20.1% and reduced KL divergence by 23.2% relative to harmonized geospatial covariates under geographically structured validation (Zhang et al., 3 May 2026). The gain was largest in larger or less-developed administrative areas where conventional settlement indicators were weak. However, the advantage reversed after aggregating United States ZCTA-level embeddings to county scale: conventional covariates achieved median $u(t,b)$6, compared with 0.658 for PDFM. This scale coupling demonstrates that embeddings are not automatically invariant under change of support.

Public-health applications extend PDFM to vaccination, cardiovascular mortality, dengue, cholera, and postpartum-depression risk. In a U.S.–Canada measles–mumps–rubella vaccination task, adding distance-weighted Canadian context increased border-county $u(t,b)$7 from 0.159 to 0.216, with an empirical permutation-test $u(t,b)$8 (Muslim et al., 5 Oct 2026). For cardiovascular mortality, PDFM provided performance broadly comparable to ACS covariates in nowcasting while offering a more rapidly refreshed representation. For Mexican dengue, PDFM improved one-month probabilistic forecasts but not robustly at three- or six-month horizons. For DRC cholera, PDFM improved four-week Precision@5 from 0.3435 to 0.3906 and eight-week Precision@5 from 0.3556 to 0.4198. In postpartum-depression prediction, PDFM produced a small AUC improvement in unseen U.S. states but did not replace individual income or Medicaid information and did not close measured demographic screening gaps.

The Planetary Prediction Engine uses PDFM as a frozen feature extractor within an autonomous workflow that selects data, performs multimodal curation, applies leakage controls, searches downstream model families, and generates predictions (Ma et al., 26 Aug 2026). The strongest reported results generally arise from combined systems:

$u(t,b)$9

Consequently, improvements cannot always be attributed to PDFM independently.

Other PDFM-style systems use distinct learning objectives. CalM trains a neural quantizer and Dual-Axis Transformer solely on calcium traces. The quantizer uses reconstruction, correlation, commitment, entropy, orthogonality, and auxiliary autoregressive losses. The transformer uses causal temporal attention, bidirectional neural-axis attention, scheduled sampling, and neighborhood replacement. CalM supports autoregressive population forecasting and downstream behavior decoding (Xu et al., 3 Apr 2026). Its multi-session behavior-decoding performance improves average $u(t,a,b)$0 by 10.1% over pretrained POYO+ on held-in data and by 7.2% on held-out data.

The latent Fokker–Planck model instead learns from independent population snapshots. Its objective combines transformed snapshot likelihood, latent-model likelihood, and hyperelastic deformation energy:

$u(t,a,b)$1

The latent dynamics are Ornstein–Uhlenbeck, while a monotone Knothe–Rosenblatt transport map represents observation-space deformation (Huang et al., 24 Jul 2026). This approach is analytically tractable but does not yet provide broad pretraining, high-dimensional validation, or demonstrated cross-system transfer.

6. Limitations, controversies, and future development

A central misconception is that PDFM is necessarily a demographic simulator. The operational geospatial model is primarily a place encoder. It does not, by itself, represent births, deaths, migration flows, age transitions, individual trajectories, or mechanistic population balance. “Dynamics” often refers to the temporal availability or downstream use of geographic representations rather than to an internally generated population process (Bell et al., 21 Oct 2025).

A second misconception is that a strong predictive association is causal. Search behavior, maps, busyness, weather, air quality, and embeddings can predict health or socioeconomic outcomes without identifying causal mechanisms. PDFM does not establish that changing an embedding-associated factor changes the target. Public-health deployments therefore require intervention studies, local validation, uncertainty monitoring, and governance.

A third issue concerns transfer across space and scale. Embeddings may encode geographic and graph-specific relationships that are meaningful at their native support but unstable after aggregation. The population-estimation study demonstrates this limitation explicitly (Zhang et al., 3 May 2026). Administrative boundaries also introduce ecological fallacy, modifiable areal-unit effects, and support mismatch.

Data sparsity and digital-coverage bias are persistent constraints. Search, Maps, busyness, and mobile-phone signals may underrepresent populations with limited internet access, smartphone use, language coverage, or digital participation. In Malawi, poor performance for TB and malnutrition was associated with small usable samples, missingness, structural zeros, and limited denominators (Metz et al., 29 Oct 2025). In the DRC, the standard self-supervised encoder could not be trained because of sparse digital data, requiring a pooled and PCA-based low-resource variant (Muslim et al., 5 Oct 2026).

Temporal limitations are also significant. Most geospatial analyses use static or single-month embeddings. Historical embeddings are not always retained, and the original data sources may be temporally misaligned. Monthly representations improve nowcasting, but long-horizon demographic forecasting, cohort aging, migration, and rapid displacement are not established capabilities.

Mechanistic PDFMs have complementary limitations. The PAT model omits migration, sex structure, socioeconomic heterogeneity, resource constraints, and changing demographic behavior. The biological-age model uses constant event rates, deterministic multiplicative jumps, a scalar biological-age coordinate, and no mortality or migration in its base version (Demongeot et al., 2023). The melanoma model uses a domain-specific MITF rheostat, sinusoidal proliferation window, and density-based stress proxy. The reaction-network framework faces structural non-identifiability, because multiple microscopic networks can produce the same deterministic dynamics (Araujo, 2023). The mean-field chiplet model is specialized to two-dimensional controlled electrostatic interactions and does not constitute a general-purpose foundation model (Nodozi et al., 2023).

A comprehensive PDFM would need to combine the strengths of these approaches while addressing their limitations. Its architecture would likely include:

• multimodal encoders for spatial, demographic, behavioral, environmental, biological, and event data;
• graph or relational operators for proximity, mobility, interaction, and administrative structure;
• temporal state-space or neural-operator components;
• density-, measure-, or distribution-valued population states;
• mechanistic modules for reaction, transport, diffusion, aging, phenotype switching, or demographic balance;
• probabilistic decoders for counts, rates, trajectories, and spatial distributions;
• hierarchical reconciliation across individuals, local units, regions, and countries;
• explicit conditioning on interventions, controls, and environmental covariates;
• uncertainty decomposition and calibrated predictive distributions;
• privacy, fairness, provenance, and abstention mechanisms;
• evaluation under spatial, temporal, demographic, and system-level distribution shift.

The most general lesson is that population persistence and evolution cannot be inferred from marginal averages alone. Refuge geometry and spatial memory, age structure and replacement, biological-age tails, phenotype selection and hysteresis, reaction mechanisms and stochastic fluctuations, geographic graph context, and latent distributional transport can all alter population outcomes. A mature PDFM would therefore need to learn not only representations of populations, but also the operators that transform population states under time, space, uncertainty, and intervention.