My research develops methods for the modeling, simulation, analysis, and optimization of supply chain networks (SCNs). The work advances along three connected themes.
Modeling and Simulation of Supply Chain (SC) Problems. Supply chains are complex, stochastic systems with intricate behaviors. Traditional analytical methods often fall short in capturing this complexity. In contrast, simulation-based methods can effectively represent such intricate dynamics. To facilitate this approach, we developed SupplyNetPy, a simulation library that currently models inventory dynamics and supply chain disruptions.
Simulation-Based Analysis and Optimization. Building on this foundation, I focus on analyzing and optimizing supply chains (SCs) through simulation. SupplyNetPy is a modular, component-based open-source library that we designed and developed specifically for modeling and simulating various supply chain problems. With this library, we can model and simulate diverse supply chain networks (SCNs) with multiple configurations, generating the performance data necessary for effective analysis and optimization.
Metamodeling with Graph Neural Networks (GNN). Optimizing an SC through simulation is costly, because every candidate design must be simulated many times under different configurations. A metamodel helps reduce this cost by learning to approximate the simulation, allowing it to be used within the optimization loop. Supply chain networks (SCNs) are inherently graph structures. Traditional metamodels, such as Gaussian process regression and standard neural networks, assume that the SCN structure is fixed, which limits their application to arbitrary SCNs. In contrast, a GNN learns from the graph structure along with the parameters of the nodes and edges, enabling the joint optimization of the network structure and its parameters. My research focuses on developing GNN metamodels for the simulation-based analysis and optimization of supply chain problems.
Sensitivity Analysis. The central goal is to exploit the differentiability of GNNs for gradient-based design-space exploration and sensitivity analysis of SCNs, and in turn for the joint optimization of their parameters and structure.
A metamodel can enhance the efficiency of SBO in supply chains. This study investigates different metamodels, specifically Gaussian Process Regression (GPR) and Neural Networks (NN), to identify optimal inventory parameter values that maximize profit and minimize costs in a five-node supply chain network (SCN). The process begins by generating simulation data and conducting an exhaustive grid search for SBO across various configurations. This approach not only identifies the optimal inventory parameters but also provides a reference solution against which the performance of metamodel-based optimization can be evaluated.
For each metamodel GPR and NN, we train the metamodel, tune its hyperparameters, assess its goodness of fit, apply local optimizers to locate the optimum, and compare that optimum against the exhaustive SBO solution. Through this comparison, the study addresses four design decisions in metamodel-based optimization:
This work is described in the paper "An Open Tool-set for Simulation, Design-space Exploration and Optimization of SCs and Inventory Problems" presented at SIMULTECH 2023, Rome.
This study lays the foundation for the work that follows. It also exposes a key limitation of traditional metamodels: they cannot account for SCN structure (topology). Each metamodel is therefore tied to a single fixed network and cannot predict the performance of arbitrary SCs. This limitation motivates the graph-based approach pursued in the GNN Metamodeling project below.
Metamodel-based approach for SBO of SCs requires training data from a simulation model. This drove us to investigate existing tools and frameworks for supply chain simulation modeling. We reviewed recent studies that model and simulate various aspects of supply chains, including inventory, logistics, resilience, and demand forecasting, for SC analysis and optimization. We also reviewed the tools/frameworks these studies use to model various aspects of SCs. The review reveals a dearth of well-maintained, open-source tools with supply chain-specific components, which motivated the development of SupplyNetPy.
This review builds upon the InventOpt work and is currently being prepared for submission. The review artifacts, including the identified node, edge, and network parameters, as well as the performance metrics, are already available in the InventOpt repository.
Visit the GitHub repo to learn more.
SupplyNetPy was developed to fill a gap identified by the literature review that continues the InventOpt work, namely the scarcity of open-source tools dedicated to SC simulation. It provides configurable, modular components, including product, inventory, suppliers, manufacturers, distributors, and demand points, from which custom SCN architectures can be assembled. Configurable inventory and replenishment policies provide flexible control over stock, and the library also supports modeling product perishability and SC disruptions.
A user can create nodes (suppliers, manufacturers, distributors, retailers, and demand), link them to create an SCN, run simulations, and evaluate key performance metrics such as throughput and profit. The SCN can be easily reconfigured and rerun for what-if analysis or simulation-based optimization.
SupplyNetPy is described in the paper "SupplyNetPy: An Open-Source Python Library for High-Fidelity Modeling and Simulation of Arbitrary Supply Chain and Inventory Networks" (to appear, WSC 2026).
For more details, visit the GitHub repo or the documentation.
Where classical metamodels are tied to a single fixed network, this work investigates Graph Neural Networks (GNNs) as structure-aware metamodels for SC performance prediction. In this work, we explored different GNN architectures to learn SCN performance from its structure and parameter configurations. We designed a custom message-passing GNN, edge-conditioned and node-type-aware, that predicts five performance metrics spanning both the node and network level (inventory carrying cost, inventory replenishment cost, transport cost, revenue, and profit) at R2 ≈ 0.99. A single trained model generalizes to networks more than 20× larger than those seen during training.
The study is accompanied by a public dataset of 100,000 programmatically generated SC graphs, built with SupplyNetPy, and by an analysis of the trade-off between accuracy and computational cost that shows a single GNN forward pass to be orders of magnitude faster than a full simulation. Unlike classical metamodels, a differentiable GNN processes node parameters, edge parameters, and graph structure together, which opens a path to the joint optimization of network structure and parameters.
This work is described in the paper "On the Potential of Graph Neural Networks as Metamodels for Supply Chain Optimization: Dataset, Architectures, and Directions" (to appear, WSC 2026).
The work continues along a single arc toward differentiable, structure-aware SC optimization, in three directions: