Introduction

Modern global supply chains have outgrown linear models. In Creating a Simulation of an Entire International Supply Chain Using Multi-Layer Perceptrons, Adarsh Keshri and the Cosminder team detail a novel framework that combines MLPs (multi-layer perceptrons) with simulation science to replicate, analyze, and optimize international logistics. This blog summarizes the key findings and implications of the 200+ page research monograph.

Motivation Behind the Research

Global supply chains are fragile, complex, and data-rich. Classical models fall short in handling non-linearity, real-time signals, and high-dimensional interactions. This study introduces MLPs to:

• Forecast demand, lead time, disruptions, and costs with greater precision.
• Adapt to policy shifts, economic changes, and real-time data feeds.
• Build digital twins of supply chains that simulate not just events—but reactions, feedback, and policy responses.

Key Components of the Simulation Framework

• Discrete-Event Simulation (DES): For modeling events like arrivals, processing, or shipments.
• Agent-Based Modeling (ABM): For dynamic agent behaviors, like retailer decisions or supplier negotiations.
• System Dynamics (SD): For macro-level flows and feedback loops over time.
• MLP Integration: Used across all layers to forecast variables (e.g., delivery delay probabilities, demand spikes).

MLP Model Architecture and Implementation

• Deep MLPs with 3–5 layers were found effective when predicting multivariate supply chain outcomes.
• Input features included demand history, port congestion, tariffs, economic indices, etc.
• Output types varied: regression for costs and times, classification for disruption likelihood.
• Employed embedding layers to handle high-cardinality data (e.g., port IDs, SKU codes).
• Dropout, batch normalization, and Adam optimizer helped stabilize training.

Real-World Case Studies

1. Electronics Supply Chain:

   • Predicted lead-time variance due to supplier delays in Southeast Asia.
   • MLP forecasts helped re-optimize transport routes and inventory buffers.

2. Pharmaceutical Supply Chain:

   • Used sensor-based cold chain data.
   • MLPs predicted spoilage risks, prompting adjusted routing and timing.

Results and Impacts

• Forecasting Accuracy: MAPE reduced by 18–25% compared to ARIMA and random forests.
• Cost Savings: Simulation-informed policy changes cut logistics costs by 12% in trials.
• Responsiveness: Integrated MLP models allowed simulations to adapt in near real-time to IoT data, tariff shifts, or fuel surges.

Research Vision and Future Scope

Cosminder’s framework envisions living simulations—self-updating models that not only forecast but learn and prescribe decisions in real time. Planned extensions include:

• Integration of transformer-based architectures.
• Expansion to carbon footprint simulations for sustainability metrics.
• Inter-firm data sharing protocols for collaborative simulations.

Final Thoughts

This research redefines how enterprises can model, test, and improve their supply chains using modern AI. With MLPs at the core, simulations become more than retrospective analyses—they become real-time strategy tools.

Disclaimer: This summary is based on "Creating a Simulation of an Entire International Supply Chain Using Multi-Layer Perceptrons" by Adarsh Keshri, 2025. For detailed mathematical appendices and code, contact support@cosminder.com.

Sources:

• Internal research by Cosminder Solutions
• A Cosminder Research Book