Supercharge Your Spatial Models with Google PDFM in CARTO
We recently explored how foundation models are reshaping geospatial analytics, and shared our vision for building new, collaborative models focused on population dynamics and spatial demographics. Foundation models open the door to broader, faster, and more consistent modeling of how people move, live, and interact with physical space. But most of these models remain in the hands of researchers and require advanced technical skills to use. Our goal is to change that.
To make this vision become a reality, we’re excited to announce the release of a new CARTO Workflows package that allows users to import and integrate Google’s Population Dynamics Foundation Model (PDFM) embeddings directly into their spatial analyses. No coding required. Just drag, drop, and connect — like any other node in Workflows.
Why does this matter? Because embeddings offer compact, high-dimensional representations of spatial areas that capture behavioral, demographic, and functional patterns in a single vector. They're a new kind of geospatial building block.
But how useful are they in practice? To find out, we ran two real-world examples using Iowa’s open Liquor Sales data, comparing model performance across traditional sociodemographic data and these new embeddings.
Understanding total sales by geographic unit is a key task across many industries. Whether it’s a CPG company estimating demand for product categories, a retail chain planning store expansion, or a public agency allocating licenses or services, spatial sales prediction informs strategic decisions.
While many organizations rely on detailed demographic data — like income, education, or population density — these datasets can be outdated, difficult to integrate, or unavailable at the necessary spatial resolution. PDFM embeddings offer an alternative: pretrained, high-dimensional spatial representations that encode behavioral and functional patterns without requiring manual feature engineering.
To evaluate how well these embeddings perform in a common business scenario, we used Iowa's open Liquor Sales dataset to predict total annual liquor sales by ZIP Code Tabulation Area (ZCTA) across the state.
We compared three model setups using CARTO Workflows:
- ACS Demographic Data only
Classic predictors, including income, education, population, etc.
- Google’s PDFM Embeddings only
No handcrafted features — just the 330-dimension vectors representing each ZCTA.
- Combined ACS + PDFM Embeddings
Fusing demographic signals with learned behavioral patterns.
The results:
The combined model delivered the most accurate results, outperforming both the ACS-only and PDFM-only approaches. Notably:
- PDFM embeddings alone outperformed ACS demographics, suggesting they capture important behavioral and functional patterns missed by traditional demographic variables.
- ACS data still contributed meaningful signals, particularly variables related to renter population and employment in arts, entertainment, recreation, accommodation, and food services.
- The best results came from combining both: a model that fuses curated demographic data with the learned spatial context encoded in PDFM embeddings. Interestingly, while the top four most predictive features came from ACS, the presence of the embeddings improved model robustness and generalization.
These results suggest that foundation model embeddings can enhance traditional models by capturing latent spatial patterns that are often missing from conventional data sources. They're particularly useful when demographic data is missing, low-quality, or needs to be augmented.
.png)
While total sales prediction is useful for understanding the broader market, CPG companies are often focused on the performance of their own products. This second use case looks at predicting sales of a single liquor product (Hawkeye Vodka) across ZIP Code Tabulation Areas (ZCTAs) in Iowa.
Using the same datasets and modeling approach as in the previous example, we tested how different input types performed for this more granular, brand-level task.
Interestingly, the PDFM embeddings slightly outperformed the other two models reinforcing a key insight: different prediction tasks benefit from different data types. Foundation model embeddings are especially powerful when modeling specific consumer behaviors or product-level demand, making them a valuable addition to any spatial analytics workflow.
With Google’s PDFM embeddings now available in CARTO Workflows, enriching your spatial models with high-dimensional, pretrained geospatial features is easier than ever. Analysts and data teams can easily plug these powerful spatial representations into predictive models — no coding required. Whether you're exploring retail demand, infrastructure needs, or market segmentation, this new building block opens up more flexible and scalable modeling.
As foundation models continue to advance, integrations like this will be essential in making spatial AI more accessible, scalable, and impactful.
.png)
