with Dr. Song Gao, Michelle Miller and Dr. Alfonso Morales,
University of Wisconsin Madison, May 21, 2026

The University of Wisconsin-Madison hosted a collaborative web presentation detailing
the development of artificial intelligence (AI) and cyberinfrastructure tools designed to
democratize food networks. Led by Alfonso Morales, a Vilas Distinguished Achievement
Professor in the Department of Planning and Landscape Architecture, the presentation
highlighted interdisciplinary efforts to model and support food supply chains across
national, regional, and local scales. The Wisconsin component of the ICICLE project
also featured work from Dr. Song Gao from the Department of Geography and Michelle
Miller from the Center for Integrated Agricultural Systems. The team works closely with
Dr. Erman Ayday and his research team at Case Western University.

The primary motivation behind this research is to counter the analytical advantages held
by national firms by making advanced algorithmic tools available to small-scale
enterprises, local governments, and non-profit organizations. Food networks
encompass highly complex activities, including production, processing, distribution, and
consumption. Historically, modeling these systems has been hindered by significant
data challenges, such as datasets fragmented across disparate spatial-temporal scales,
extreme data sparsity, a lack of predictive planning tools, and persistent privacy barriers
to data sharing. By implementing a multi-scaled reference architecture, the ICICLE
project delivers software solutions that empower local and regional stakeholders to
collaborate, preserve data privacy, and execute predictive simulations on the fly.

The National Scale: Geospatial AI and Global Resiliency

Dr. Song Gao manages the national-scale research efforts within ICICLE, leveraging his
background in Geographic Information Systems (GIS) and Geospatial AI to resolve data
friction on a macro level. To analyze the movement of food across the United States,
Gao’s team utilizes the federal Freight Analysis Framework (FAF) zone-level dataset,
which tracks food flows across 132 national food zones across seven standardized
transportation commodity categories.

To conceptualize and evaluate macro-network structural health, the team built several
crucial software structures:

• Geospatial Knowledge Graph: This framework maps national entities, food
  categories, and logistical relationships to evaluate food network resiliency on the
  fly. Resiliency is determined by whether a specific geographic region maintains a
  diversity of suppliers, a wide variety of available food commodity types, and
  optimized, short-distance transportation routes.
• GM Food Flow Modeler: Building upon the legacy data-driven frameworks
  established by pioneers like Dr. Deniz Karacok—which originally relied on classic
  ridge regression techniques—Gao’s team integrated advanced deep learning
  methodologies. They constructed a highly scalable food flow model covering over
  9 million links across more than 3,000 U.S. counties using Graph Neural
  Networks (GNNs). This GNN approach specifically overcomes severe data
  sparsity issues to deliver robust performance metrics.
• Hurdle-Based Geo-Architecture: The GNN model relies on a coupled, two-part
  hurdle architecture. The first module handles link classification to determine
  whether a physical food flow exists between two specific counties. The second
  module runs a flow regression to estimate the exact volume or value of that
  specific food category moving between those counties. Because ground-truth
  data at the county level is heavily aggregated by the U.S. Census, the team
  evaluates spatial interactions using the Common Part of Commuters (CPC)
  metric instead of traditional $R^2$ values, which can fail when evaluating highly
  sparse matrices.

The practical application of this macro-level modeling is made available via the
standalone GM Food Flow portal as well as the integrated ICICLE AI Service Portal,
which hosts domain-specific AI services spanning digital agriculture, animal ecology,
and food systems. A core capability of this platform is its interactive Scenario
Simulation Mode. Users can ask “what-if” questions regarding macroeconomic shifts.
For example, a user can simulate changes in agricultural production variables (such as
a reduction in vegetables or soybeans) within a massive exporting center like Los
Angeles County, California, and immediately visualize the cascading, downstream
impacts on specific calorie availability across destination counties nationwide.

The Regional Scale: Supply Chain Optimization and Logistics

Michelle Miller addresses food systems at the regional scale, focusing on translating
data science into actionable insights for small scale wholesale agricultural practitioners
and regional food distributors. Miller points out that companies like Amazon use in-
house digital business ecosystems to dominate fresh food markets, while 90% of
Fortune 500 firms successfully employ third-party logistics (3PL) providers to optimize
their operations. For independent regional food networks to compete, lower their costs,
and operate within fair markets, they require collaborative digital and physical
infrastructures. These include shared cold chain resources, cooperative route
management, and data-driven modeling.

A primary challenge in regional food logistics is managing the operational divide
between Full Truckload (FTL) and Less-Than-Truckload (LTL) shipping. FTL
movements are simple and cost-effective because an entire truck is filled with a single
product moving from one origin to one destination. Conversely, regional fresh food
systems depend heavily on LTL multi-drop deliveries, which introduce severe
operational costs and routing complexities. While European logistics frameworks
subsidize and support chilled part-load freight movements, U.S. infrastructure favors full
refrigerated trailers (“reefers”), forcing small regional distributors to optimize complex
variables—product dimensions, weight, distance, temperature control, vehicle assets,
driver availability, seasonality, and varying client constraints—manually and often on the
fly, as conditions and resources change.

To modernize these operations, Miller’s team is working on specialized digital tools:

• Tabular Data Synthesizer: most small businesses rely entirely on tabular data,
  this software synthesizes tabular logistics data for ingestion into AI-powered
  optimization tools.
• Synthetic Data Generator: AI models require large amounts of data. This AI-
  driven tool generates synthetic location and vehicle data, giving software
  developers the dummy datasets required to safely test and iterate logistics
  software without risking sensitive business details.
• Vehicle Routing Model: Utilizing proprietary operations data from the
  Wisconsin Food Hub Cooperative, the team is developing a prototype vehicle
  routing model tailored to the geographical layout, seasonality and scale of the
  Upper Midwest.

Recognizing that 80% of transportation costs are directly tied to infrastructure location,
this routing framework is built to optimize both “first-mile” and “last-mile” logistics. First
mile logistics involves moving products from individual farms to a central facility. If
optimizing both location and routing, this has been shown to yield a 9% reduction in
operations costs. Last-mile logistics to rural grocers, institutions and Tribal communities
are penalized by urban-centric distribution models. Ultimately, this software will be
delivered as a vetted, open-source routing model that small-to-medium wholesale
businesses can run locally on their own servers without relying on expensive,
centralized data centers or cloud computing.

The Local Scale: Granular Access, Workflows, and Data Privacy

Dr. Alfonso Morales focuses on the local scale, emphasizing workflow integration
across business silos and designing software around the explicit needs of municipal
governments and non-profit organizations. Operating out of the Kauffman Lab, Morales
employs an applied, community-partnered methodology where practitioners function as
co-principal investigators whose real-world expertise directly drives software
development.

Morales highlights three central software systems active at the local scale:

• Farmtofacts.org: Brought into the ICICLE framework as a mature data-
  gathering platform (currently in its 2.x release), Farm to Facts captures highly
  granular, vendor-level data from farmers’ markets. It helps markets across the
  United States and Canada track local economic impacts, social metrics, and vital
  ecosystem services.
• FEAST (Food Equity Access Simulation Technology): Also known as the
  Food Access Simulator, FEAST is an agent-based modeling tool designed for
  decision support. Deployed via high-performance computing (HPC) infrastructure
  at the Texas Advanced Computing Center (TACC), FEAST utilizes Kubernetes
  pods called TAPIS to execute complex computational models rapidly. The
  software provides an interactive mapping interface that visualizes data down to
  the census tract, block group, and individual building levels. Planners can overlay
  the spatial coordinates of grocery stores, supermarkets, farmers’ markets, and
  SNAP-authorized retailers alongside critical environmental vulnerabilities like
  regional floodplains.
• Privacy-Preserving Web Portal: To allow stakeholders to integrate proprietary
  business logistics or localized data safely into tools like FEAST without exposing
  trade secrets, Dr. Erman Ayday’s team at Case Western University developed a
  privacy-preserving web portal. Powered by advanced genomic privacy principles,
  this cyberinfrastructure allows independent entities to securely mask, share, and
  co-model sensitive data together, ensuring community collaborations can
  progress “at the speed of trust”.

FEAST features a powerful Scenario Mode that enables local planners and hunger-
relief organizations to simulate structural interventions on the fly—such as measuring
the impact of adding or removing a supermarket or calculating the optimized placement
of a mobile food market by organizations like Feeding America. Brown County,
Wisconsin, served as the first paying client for FEAST, using the software’s scenario
engine to coordinate emergency food distribution responses to localized flooding.
https://foodaccesssimulator.com/counties/brown-county

Multi-Scalar Integration and Long-Term Vision

The long-term vision of the Wisconsin ICICLE team centers on workflow integration
across all three geographic scales, joining them into an interconnected Data and Model
Commons tied to a unified reference architecture. Under this framework, a highly
localized disruption modeled in FEAST can be contextualized against regional trucking
routes optimized by the Vehicle Routing Model, which in turn informs macro-level supply
changes mapped by the national GM Food Flow modeler.

To make these tools accessible to resource-constrained organizations like local food
banks or small agricultural co-ops, the team utilizes a “stone soup” approach to
software development. Under this philosophy, any specific feature or analytical tool
engineered for a paying client (such as a specific municipality) is immediately
modularized and made portable as an open-source asset for all other users in the
network. This collective scaling drives down customized development costs
dramatically—bringing a comprehensive, county-level simulation down to an accessible
$30,000 range—and successfully democratizes algorithmic power to protect food
security at every scale.

Questions from Kushank Bajaj, collaborator with the Greater Vancouver Food Bank

Question 1 (On Reproducibility and Long-Term Vision): Who is expected to
run these models in the long term, and will they be easily reproducible for small
enterprises?
o Answer by Michelle Miller: For the regional routing models, the primary
users will be regional distributors themselves, such as the Wisconsin Food
Hub Cooperative. The Food Hub operates as a centralized distribution hub
with “spokes” reaching out to smaller independent distributors across the
state—a collaborative, multi-tenant model that has proven highly
successful in France.
o Answer by Alfonso Morales: For the local software, the platforms are built
explicitly for practitioners, local governments, and non-profits to utilize as
decision-support systems. Farm to Facts already features a mature
Canadian instance operating with the Farmers’ Markets of Nova Scotia,
and the FEAST software can be adapted similarly to meet Canadian
municipal needs. The long-term plan is to fully integrate these distinct
software tools so that local food entrepreneurship data directly informs
broader food security models.

Question 2 (On Financial Accessibility for Resource-Constrained Food
Banks): How can a small co-op or a food bank with very few financial resources
mobilize and access these models?
o Answer by Alfonso Morales: Financial barriers are minimized through local
philanthropic partnerships and the team’s “stone soup” software model.
Because every new feature developed for a client becomes a portable
asset shared across the open-source network, baseline costs drop
significantly. For example, during a recent presentation to the City of
Milwaukee, the team estimated that a comprehensive simulation model for
a major area like Milwaukee County would cost in the accessible range of
$30,000, a fee often easily covered by local philanthropists.

Question 3 (On Methodological Differences in Food Flow Modeling): How
does the national-scale model differ from other prominent food flow frameworks,
such as the work of Dr. Megan Konar?
o Answer by Song Gao: The defining difference is ICICLE’s deep focus on
real-time scenario simulation. Rather than running a static, one-time
calculation of historical data, the GM Food Flow platform allows users to
manipulate variables interactively and run active simulations on the fly to
see immediate network impacts. Gao and Miller both agreed that hosting a
dedicated national food flow workshop or symposium would be an
excellent step to bring these different research groups together.

Questions from Kamran Zendehdel, USDA Agricultural Marketing Service, Local & Regional Food Division

Question 1 (On End-Users): Who exactly will be the day-to-day user of the
regional transportation and logistics tools being developed?
o Answer by Michelle Miller: The users are small, regional distribution
companies. For instance, the Wisconsin Food Hub Cooperative manages
fewer than 20 routes, sourcing local produce within a 200-mile radius.
They consolidate goods at their warehouse and distribute them to school
districts, independent grocers, larger clients like Kroger, and remote Tribal
communities (managing 11 Tribal drop-offs every other week). These
small hubs desperately need this software because managing these
complex, fragmented routes manually is highly inefficient.

Question 2 (On Scalability for Small-to-Mid-Sized Producers): Is it possible
to extend these regional routing tools so that individual small and mid-sized
producers can use them to determine their best method of distribution?
o Answer by Michelle Miller: Yes, the underlying model is designed to be
completely generic and entirely open source. While the initial training data
sourced from the Wisconsin Food Hub Co-op is proprietary, the
algorithmic model itself is not. This allows any independent small-to-mid-
sized business or regional hub to plug their own private operational data
into the model to manage and optimize their local distribution networks
independently.

This webinar series is made possible by the ICICLE – Intelligent CI with Computational
Learning in the Environment – U.S. National Science Foundation AI Center under OAC-
2112606 and organized by the UW-CALS Center for Integrated Agricultural Systems,
and with support from several UW partners: the Grainger Center for Supply Chain
Management, the Kaufman Lab, Center for Sustainability and the Global Environment
(SAGE), Organic Collaborative, Center for Cooperatives, and the Food Studies
Network.