Problem:
Amazon needed to see how pallet loading decisions affected both local operations and the wider network. A rule that worked for one lane could increase trailer demand elsewhere.
Solution:
The team used network science simulation with a 3D bin packing heuristic to test pallet loading scenarios. AnyLogic supported GIS calculations, algorithm testing, A/B testing, and visualization.
Results:
- Modeled trailers with up to 52 pallets.
- Found that 100% industry-standard palletization would add hundreds of trailers.
- Determined that excluding the top 3 highest-volume arcs reduced hundreds of trailers to only a handful.
- Identified a strategy to improve pallet utilization by up to 85%.
Introduction: using simulation to improve operational efficiency
Amazon’s Worldwide Design Engineering team works across a complex fulfillment network, where products move from suppliers to customers through multiple facilities and transportation stages.
To improve operational efficiency, the team uses a science-driven workflow that combines strategy, validation, process and product development, and operational standards. This workflow brings together mathematical modeling, data analysis, simulation, emulation, artificial intelligence, machine learning, and visualization.
Check out other Amazon case studies:
Simulation-Driven Solution for Fulfillment Logistics Evaluation
Simulation for Transportation Network Optimization via Truck Yard Revision
Problem: local efficiency was not enough
At the local level, the team had to consider throughput, process efficiency, and on-time order fulfillment. At the network level, they also needed to account for employee safety, service quality, cost efficiency, sustainability, and adaptability to future market trends.
These objectives could conflict: a decision that improved speed in one part of the operation could increase transportation demand elsewhere. A standalone view was not enough. Amazon needed models that could track unit-level, facility-level, and network-level metrics, including miles traveled per unit, delivery time, throughput efficiency, and daily trailer counts.
In up to 70% of cases, Amazon used to rely on a full (fluid) palletization strategy. The team wanted to test other possible ways of palletizing risk-free.
The team wanted to understand which changes could improve operational efficiency across the entire network.
Solution: a network science simulation approach
At earlier stages, Amazon developed a network science simulation approach that connected micro-level operational behavior with macro-level network behavior. This allowed the team to analyze what happened inside facilities and how those decisions affected transportation flows across the network.
The Amazon network model (click to play)
The simulation environment tracked several levels of metrics:
- Unit-level: miles traveled per unit and delivery time.
- Facility-level: throughput and efficiency.
- Network-level: number of trailers running day to day.
Pallet loading and outbound dock configuration
For this case, Amazon combined network science simulation with a 3D bin packing heuristic based on the Extreme Points Method. This helped the team evaluate how cases with different dimensions could be placed on pallets and how different pallet loading strategies would affect trailer utilization, pallet utilization, and trailer count.
The following pallet loading strategies were considered:
- Fluid (case only): when individual cartons are loaded directly into the container without pallets.
- Hybrid: a combination of palletized and loose-loaded freight.
- Pallet: products are fully palletized before loading (considered an industry standard).
- Fluid (tote only): reusable totes/bins are loaded loose.
Different modes of outbound loading that Amazon uses
The 3D bin packing model was simulated using two weeks of historical data for all inbound nodes. It assumed random sampling of package dimensions, including length, breadth, and height, at the origin-destination pair level. The model also assumed that current fluid-loaded trailers could be converted into palletized loads.
To test different pallet loading conditions, the team used three open pallet positions per trailer as the base case. Then they varied this number from one to six for sensitivity analysis.
The model also used several fixed operational assumptions:
- tote dimension: 2.3 cubic feet
- maximum pallets per trailer: 52
- UPP dimensions: 64 cubic feet, based on a 4x4x4 ft unit
The heuristic accounted for box placement, non-overlap, physical support, bin closing conditions, and box orientation. This helped estimate how palletized loading would affect pallet utilization, trailer utilization, and trailer count across the network.
Results: insights that helped improve operational efficiency
The simulation helped Amazon evaluate outbound loading strategies for both the current and future network setup. Instead of applying the same pallet loading rule to every lane, the team could make lane-level recommendations based on actual volume flows and network dynamics.
The model showed that converting 100% of fluid-loaded trailers to palletized trailers would require hundreds of extra trailers, while excluding only the top three highest-volume arcs reduced the increase to just a handful.
This proved that a hybrid pallet loading setup could improve operational efficiency better than a blanket conversion strategy. The study also showed that pallet utilization depends strongly on the number of open pallet positions available for each destination. By controlling these positions more tactically, Amazon could improve pallet utilization by up to 85%, making pallet loading a practical lever for balancing process efficiency, trailer utilization, and transportation impact.
The case study was presented by Siva Veluchamy, Prasad Rao, and Mahek Chheda from Amazon at the AnyLogic Conference 2025.
The slides are available as a PDF.