Supply chain network design models are the most widely leveraged strategic models in the supply chain domain. They are also the most widely challenged models!

Classic optimization models have been leveraged in the operations and supply chain world for over half a century. I do not need to highlight that the world has changed drastically. And so has the world of operations and supply chains.

Supply chain network design models are strategic models. They are generally focused on designing and optimizing nodes and flows of a supply chain. Due to this very reason, these models are limited by their level of granularity. And this is where the challenges surface.

If you have been in roles where you presented findings from a model you developed to senior management, you know where I am coming from. While the model does not factor in hundreds of granular aspects, those in the room will have questions about the same constraint the model cannot factor in its optimization.

Off-the-shelf supply chain network design software companies have tried to address this. You can use some development work to design your customized constraints. And if you are writing your code from scratch, you can insert every constraint. But then your run times become unrealistic if your model has many products, sites, etc.

I have suggested leveraging deep learning in tandem with optimization models in articles like this one for tactical modeling. But supply chain design was not an area I contemplated or experimented with.

But this week, I read a research paper titled “Selecting the Location of apparel manufacturing plants using neural networks” by Zeng et al. Zeng and others propose a methodology to leverage NN to select optimal manufacturing locations from a list of manufacturing locations. The model uses classification to categorize the sites based on criteria like:

• Land and construction
• Labor
• Transportation
• Proximity to markets
• Utilities and real estate
• Country Status
• Government policies
• Competition
• Machinery and services
• Community conditions

My first thought was that you could also incorporate all these criteria in an optimization model. And since these models are not models that you would run in real-time or frequently (like tactical or operational models), there might not be a benefit to invest in developing NN models for optimal site selection.

But only if we want to constrain the model to site selection!

Supply chain network design goes much beyond site selection. For example, you may only aim to optimize flows in some scenarios. Any optimization has implications that percolate beyond the strategic level, seeping down to tactical operations. Incorporating all these aspects in a network model has been difficult, if not impossible.

And that is where we can extrapolate the approach Zeng and others have proposed. And go beyond classification.

We can leverage ANN to incorporate the impact of granular aspects of network design changes. Tweaking the approach of suitability index score leveraged in the paper, we can build models that will incorporate granular aspects down to shop floors. The slotting of products on the warehouse floors is an example of a granularity level that can be factored in. Impact on pick time is another example.

This can be a foundation for an algorithm that can one day be leveraged to plan the supply chain across all levels of planning semi-autonomously, from strategic to tactical.

But this capability can also be extended beyond the supply chain into the marketing world. You can extrapolate this approach to tie various levels of marketing planning into one thread.

The time may be ripe to change the methodology and perception of supply chain network design.