This work is the fruit of a partnership with Renault.
In a context of growing decarbonization effort in society, the automotive industry will have to undergo major transformations.
These historical shifts put forward the challenge to design an efficient and resilient supply chain in the long run.
One of the facets of this problem is to design the supply chain chain network.
This study adresses one facet of this general problem : designing the supply chain network for Renault's inbound operation.
The inbound supply chain of a automotive manufacturer is designed to transport car parts from numerous different suppliers to several plants.
For Renault, this worldwide network encompasses several thousand suppliers, producing over tens of thousands types of parts, dozens of intermediate logistics platforms, and dozens of plants to deliver.
This logistics operation represents an annual expenditure of approximately one billion euros and results in tens of thousands of tons of CO2 emissions per year.
Optimizing Renault's logistics network in the long run is therefore imperative.

Developing a comprehensive network design optimization tool entails optimizing in an integrated way the transportation network and the flow of car parts within the network.
The goal of this optimization is to formulate a long-term network design decisions while providing a corresponding transportation plan to be procured from carriers.
To that purpose, we introduce the Shipper Network Design Problem, a rich variant of a multicommodity network design problem (MCNDP).
To better align with Renault's specific context and requirements, several extensions to usual network design models are considered.
The transportation model used in this work is in itself an extension of a Load Plan Design Problem.
Termed Shipper Transportation Planning Problem, it integrates routing with flexible delivery time, bin-packing consolidation and novel regularity constraints.
The design model is also tailored to the specific needs of Renault.
As it contracts with logistics providers at platforms, a node design approach is considered.
Contracts between Renault and platforms work by geographic units.
The car manufacturer partitions the basic units into geographic clusters, and contracts all the supplier-to-platform transport for a given cluster with a single platform.
Cluster need to make sense geographically, which basically means that they must be contiguous and compact.
Those considerations introduce a novel districting decision in the design model.
The computational difficulty of this special MCNDP model stems from its combinatorial nature, integrating in a single model multiple usual combinatorial problems, and the combination of a large time-expanded network, hundreds of thousands of arcs, with several millions of commodities to be routed.
Although recent solution methods from the literature scale to large networks, they don't consider districting as part of the problem and typically handle no more than than fifty thousand commodities while not considering bin-packing consolidation.

We present the modeling of this specialized network design problem and propose two tailored heuristics for solving it at an industrial scale.
The first heuristic is based on an Iterated Local Search algorithm, which combines a local search developed for the transportation subproblem with custom network perturbations based on Mixed-Integer Linear Programming relaxations to explore efficiently the solution space and achieve high-quality solutions.
The second heuristic employs a learned decomposition of the problem to significantly reduce computation time, addressing the industrial need for interactive algorithm testing and strategic decision improvement.
We conduct a data analysis of Renault's instance and establish a lower bound to evaluate algorithm performance.
Our numerical experiments demonstrate a substantial improvement over Renault's current solution.