A network optimization model in three weeks, and the four steps that got there

Adapted from the update iNDustry Labs sends the AI Executive Group, a group of manufacturing and logistics leaders across the South Bend–Elkhart region.

GenAI in the supply chain: a case study

On a recent project with a regional business, the iNDustry Labs team used generative AI to accelerate the development of a network optimization algorithm. The model weighed customer demand by geography, SKU production costs at each facility and transportation costs, with the goal of minimizing the total cost to serve customers. Analysis that would once have taken months was completed in three weeks.

The point of walking through it here is not the mathematics. It is that an industrial engineer, rather than a data scientist, can now get a credible first model on the table. The example below is de-identified and simplified. If you run a single site, substitute work centers for facilities as the nodes.

1. Describe the problem and the objective

The opening prompt does the work that a scoping conversation would normally do. It names the network, the quantities and the constraint structure:

"I am working on a supply chain network optimization project. The attached picture represents the supply chain network. We have 2 plants that produce 312 different finished goods. We also have 5 distribution centers and more than 1,000 end customers grouped by ZIP code. Once production at a plant is completed, the finished goods ship to distribution centers, which then ship to end customers, so our total cost to serve includes production, transportation from plant to distribution center, and transportation from distribution center to customer. I want an objective function that minimizes cost to serve while satisfying our demand and production capacity constraints. The demand constraints are specified per product and per end-customer ZIP code grouping."

2. Review and adjust the model

What comes back is a starting point, not an answer. On this project the output was validated and adjusted by supply chain experts and data scientists before anyone trusted it. That is the realistic division of labor: the model gets an engineer to something concrete enough to hand to a more technical colleague, which is a different and more useful thing than getting it right on the first pass.

3. Develop the script

The follow-up prompt was simply to ask for an implementation: "I want to use Python to run this optimization. Attached are my data input files. Write a script that uses these files to run the objective function above." The attached files held per-SKU production costs at each factory, both plants' total capacities, and customer demand by location. The resulting script read each file, extracted the unique plants, distribution centers, customers and products for indexing, built a minimization problem with decision variables and an objective function summing costs, then imposed four constraints: demand satisfaction at each customer, flow conservation at distribution centers, linking production to outflow at each plant, and production capacity.

Before running anything, check that the general approach suits your use case, your data size and the computing resources you actually have.

4. Visualize, then decide

After solving, the team visualized the output, both to check the results passed the smell test and to leave room for human judgment. Sankey diagrams worked well here because they show the flow of product from each plant through the distribution centers to aggregated customer regions. Models produce suggestions. People still make the decision.

Where the project pipeline stands

Conversations with regional companies this cycle surfaced 21 distinct project opportunities, comfortably beyond the capacity of a single semester. The constraint on this work is not demand from industry. It is how many student teams can be supported at once.

From the events this cycle

An Innovation Park session on attracting AI-savvy talent brought together a regional manufacturing CFO, the head of the South Bend–Elkhart Regional Partnership and Notre Dame HR experts, covering how to write job descriptions that appeal to candidates who already work with AI tools, and how AI is reshaping job descriptions and pay scales. The Q1 AI keynotes that followed paired a Mendoza College of Business session on measuring AI return on investment with a regional company's account of its own AI journey.