Implementing the GEODIS digital twin, and the technology behind it

Technical appendix: digital twin methodology

This appendix provides technical details for evaluators who want to understand our digital twin architecture, technologies, and implementation approach.

Data architecture and integration

The digital twin consolidates operational data from multiple systems to create a unified view of warehouse performance. We integrate directly with the GEODIS Warehouse Management System that controls operations, pulling order data, inventory levels, task assignments, and completion timestamps. The Labor Management System contributes worker productivity metrics, attendance records, and performance data broken down by job category.

Additional data sources include:

• Warehouse Control Systems for equipment status, conveyor throughput, and sortation metrics
• RF gun scanning data capturing picking events, pack operations, and shipping confirmations
• Locus robotics systems providing autonomous mobile robot performance data
• Bastian automation equipment feeding in print operations and material handling metrics

Each technology operates on different refresh cycles based on its purpose. Predictive planning updates daily with weekly forward-looking forecasts. Process intelligence continuously collects data with monthly analysis cycles to identify long-term patterns. Virtual testing runs on-demand when modeling scenarios. Real-time monitoring refreshes every 15 minutes for near-real-time operational visibility.

Our technical integrations use several methods. Direct SQL database connections handle GWMS and LMS data since GEODIS controls those systems. API integrations connect automation equipment and robotics platforms. Python-based ETL pipelines running on dedicated servers transform and load data, with Python Airflow jobs managing the automation workflows. Azure Synapse serves as the consolidated data warehouse. For real-time monitoring, automated processes push data to Celonis Cloud every 15 minutes.

Process intelligence: Celonis platform

Celonis analyzes warehouse management system transaction logs using an object-oriented methodology. Each order becomes a tracked object moving through defined warehouse events, with the system recording order IDs, timestamps, and activity types at each stage.

The platform monitors specific warehouse operations including induct, Locus picking, detote, packing, Pandas sorting, and shipsorter stages. This granular tracking enables precise identification of where delays occur and where processes deviate from optimal workflows.

Celonis provides two primary dashboard views that our operations teams use daily. The Process Explorer tracks orders through warehouse blocks in real time, showing quantities and average times between events. The Variant Explorer analyzes process deviations to identify where actual workflows diverge from expected patterns. Both views support filtering by order type, quantity, and other parameters to drill into specific operational scenarios.

The platform automatically identifies process variants, conformance issues, cycle time problems, and efficiency opportunities. However, this analysis generates hundreds of potential findings. A dedicated process analyst reviews the Celonis output and filters it down to the top five most actionable improvements, the changes that will deliver significant impact with reasonable implementation effort. This filtering prevents operations teams from getting overwhelmed by data while ensuring they focus on high-value optimizations.

Virtual testing: FlexSim simulation

FlexSim is sophisticated 3D simulation software that GEODIS uses to model complete warehouse operations. The digital replicas include conveyors, picking stations, pack areas, sortation equipment, and worker movement patterns.

FlexSim's inbuilt conveyor models and warehouse components simplify the creation of accurate simulations. We can test equipment performance, model order flow patterns, and predict where bottlenecks will emerge under different volume scenarios. The system enables virtual testing of equipment changes, layout modifications, and process adjustments before implementing anything in the physical warehouse.

Key capabilities include:

• Scenario testing for "what-if" analysis without disrupting live operations
• Python control interface for automated test execution and parameter adjustments
• Database connectivity for near real-time monitoring integration with operational models
• Visual representation showing simulation results through graphs and dashboards
• Performance comparison between different configurations

The system uses color-coded visualization to make capacity constraints immediately apparent. Red highlighting indicates bottlenecks where demand exceeds capacity, allowing operations teams to identify problems before they occur in actual operations.

Predictive planning: machine learning algorithms

GEODIS combines traditional formula-based workforce calculations with advanced machine learning forecasting. The system uses two primary algorithms selected for their specific strengths in warehouse operations.

Prophet Algorithm is Facebook's open-source forecasting tool that automatically detects seasonal patterns, holiday effects, and trend changes in operational data. It provides reliable results for longer-term strategic planning and handles external factors like holidays and peak periods effectively.

NHiTS (Neural Hierarchical Interpolation for Time Series) represents our more advanced forecasting approach. This neural network architecture was specifically designed for time series forecasting and demonstrates superior accuracy for warehouse operations. The algorithm detects subtle productivity patterns that other methods miss and is easier to tune for individual work categories like picking, packing, and sorting. It excels at long-term forecasts where traditional methods struggle.

Both algorithms analyze historical volume and labor data combined with external variables including holidays, peak dates, and product launch schedules. The forecasting works across multiple time horizons, providing strategic planning with 365-day forward visibility (refreshed weekly) and tactical planning at the next-day and next-hour level for immediate staffing decisions.

The system calculates labor requirements by job category, breaking down exactly how many pickers, packers, and other roles are needed. Integration with the GEODIS WMS enables continuous model updates as actual performance data flows in. This creates a self-improving system where better predictions lead to tighter labor planning and earlier identification of capacity constraints.

Implementation requirements

GEODIS manages both the warehouse operations and the underlying technology platform, which simplifies implementation considerably.

For existing GEODIS clients, implementation requires:

• Access to order forecasts and demand planning data
• Collaboration between client teams and GEODIS operations teams
• Participation in review sessions and optimization meetings as the system is refined

For companies transitioning to GEODIS from another provider, the process requires:

• Historical operational data from the previous provider (if available)
• Order forecast data and seasonal volume patterns
• Product mix and SKU information
• Service level requirements and operational priorities

GEODIS can build effective digital twins even without comprehensive historical data. The machine learning models improve rapidly as they collect actual performance information from live operations. Initial forecasts may be less precise, but accuracy increases within weeks as the system learns actual productivity patterns.

All technical infrastructure is GEODIS-owned and managed. The digital twin components run on our Azure cloud environment with built-in redundancy and security controls. GEODIS establishes the SQL connections and data integration pipelines during implementation. Client organizations don't need to modify their infrastructure or invest in additional technology.

Previously in this series: Part 1, what a digital twin is and what it delivered in our pilot. Part 2, testing changes before they happen and catching problems as they form.