Machine learning and blockchain team up to cut supply chain costs and carbon

Modern supply chains juggle three competing demands at once: keeping costs down, moving goods efficiently, and shrinking the environmental footprint of every shipment. A study published on 18 June 2026 in the open-access journal Discover Analytics argues that the tools built to handle these demands have mostly evolved in isolation, and it proposes a way to fuse them into a single decision-support pipeline. The work, led by Meisam Karami and Hamid Bouraghi, combines machine learning prediction, multi-objective optimization, and blockchain-based transaction recording into one framework intended to guide supply chain decisions in real time.

The researchers identify a persistent structural gap in the literature. Machine learning studies produce forecasts of key performance indicators but rarely embed those forecasts into operational decision loops. Blockchain research emphasizes transparency, traceability, and tamper-resistant data sharing, yet the authors note that blockchain-based systems record transactions without supporting real-time predictive optimization or decision feedback. The result, they argue, is a fragmented landscape in which predictive intelligence, optimization mechanisms, and secure data infrastructures operate separately, limiting the development of integrated, time-sensitive decision-support systems for supply chains facing demand volatility, disruptions, and sustainability constraints.

To close that gap, the team built a four-stage workflow. First, supply chain operational and sustainability data are collected and preprocessed through cleaning, feature selection, and normalization. Second, a Random Forest regression model is trained to estimate key performance indicators such as order fulfillment rate and operational efficiency. Third, the predicted indicators feed into a multi-objective optimization model formulated as a Linear Programming problem, which simultaneously minimizes operational costs and carbon emissions under supplier capacity and demand constraints. Finally, a blockchain-based transaction layer records the optimized allocations to enhance traceability, transparency, and data integrity across supplier-buyer interactions.

The empirical analysis drew on three related datasets. The foundational file contained 999 rows and 23 columns of company-specific information and performance metrics, including supplier count, inventory turnover, and order fulfillment rate. An updated version added a Total Implementation Cost column, expanding to 1,000 rows and 24 columns. The final dataset extended the record to 29 columns by incorporating environmental indicators such as carbon emissions, energy consumption, and green packaging usage, which allowed the optimization process to weigh environmental impact alongside operational and financial metrics. Preprocessing involved imputing or removing missing values, converting data types, and applying Min-Max normalization so that features measured in different units contributed comparably to the models.

The Random Forest Regressor, implemented in the scikit-learn library, builds multiple decision trees on random subsets of the data and features, then averages their outputs to produce stable predictions that capture nonlinear relationships among variables. The dataset was split with 80 percent used for training and 20 percent reserved for testing. Hyperparameters such as the number of trees, maximum tree depth, and minimum samples for splitting were tuned, and the model was evaluated with standard regression metrics. The reported performance was strong: a coefficient of determination of 0.893 and a root mean squared error of 0.503, indicating that the model explained a large share of the variance in the target indicators on the held-out data.

Feature importance analysis revealed a striking pattern. Supply Chain Risk and Supply Chain Resilience Score were the most influential predictors of order fulfillment rate, together accounting for roughly 60 percent of the model’s predictive power. Moderately important features included Supplier Relationship Score, Lead Time, Revenue Growth Rate, and Inventory Accuracy, with importance scores between 5 and 10 percent. Environmental and financial variables such as Recycling Rate, Green Packaging Usage, Use of Renewable Energy, and Cost of Goods Sold showed minimal impact on the prediction. The authors interpret this as evidence that operational stability and risk management, rather than sustainability metrics, dominate fulfillment performance in their predictive model, though they caution that environmental indicators remain important for sustainability reporting.

The optimization stage formulated a weighted linear objective that minimized a blend of total transport cost and total carbon emissions, subject to constraints on supplier capacity, buyer demand, and an emissions limit. Solved with the PuLP library using the CBC solver, the model produced supplier-buyer allocations that the authors visualized in a cost-versus-emissions scatter plot and a Pareto frontier chart. The results showed a measurable trade-off: as the cost of an allocation increased, carbon emissions tended to rise as well, meaning that minimizing costs does not always coincide with minimizing environmental impact. The objective function weighted cost minimization at 0.7, and most allocations clustered in the lower to mid-range of both cost and emissions, with a few high-cost, high-emission outliers driven by capacity limits or specific buyer needs.

Supporting analyses reinforced the optimization findings. A correlation matrix showed a strong negative relationship between lead time and order fulfillment rate, suggesting that shorter lead times align with better on-time delivery, and a perfect positive correlation between energy consumption and carbon emissions, reinforcing the case for energy efficiency initiatives. Total implementation cost showed only weak correlation with performance metrics, which the authors read as a warning that simply increasing investment does not automatically improve outcomes. Supplier count showed little correlation with performance, pointing to supplier quality over quantity as the more meaningful lever. An allocation heatmap further revealed uneven distribution patterns, with some suppliers serving many buyers while others were largely unused, likely due to high costs or emissions.

The blockchain component was implemented as a prototype rather than a fully deployed industrial network. Each transaction, including supplier-buyer movements of goods, is encapsulated in a block containing identities, quantities, price, timestamp, and a cryptographic hash, with blocks linked sequentially so records cannot be altered without modifying all subsequent blocks. Transactions are digitally signed using RSA encryption, and the ledger is stored in a distributed JSON file replicated across participating nodes. The authors state that this layer ensures optimized allocations remain auditable and verifiable across supply chain nodes, supporting traceability and data integrity, and they explicitly frame the implementation as a demonstration of how blockchain structures can complement predictive analytics and optimization rather than a production system.

The authors acknowledge several limitations. The empirical analysis relies on a specific dataset and does not benchmark against alternative machine learning or optimization models such as gradient boosting, neural networks, or heuristic approaches, which they recommend for future comparative experiments. Scalability also remains an open question, since large-scale networks with thousands of suppliers, buyers, and transactions could strain the computational performance of both the optimization problems and blockchain transaction volumes, potentially requiring distributed computing and scalable infrastructure. They position the framework as a proof-of-concept analytical architecture, and future work should evaluate it on larger datasets and real-world industrial case studies to assess robustness and deployment potential. Even so, the study offers a concrete template for how prediction, optimization, and secure transaction recording can be wired together to balance economic and environmental goals in supply chain decision-making.

Subject of Research: An integrated machine learning, optimization and blockchain framework for supply chain decision-making

Article Title: Blockchain enabled real time supply chain optimization bridging financial markets and operational efficiency

Article References: Karami, M., & Bouraghi, H. (2026). Blockchain enabled real time supply chain optimization bridging financial markets and operational efficiency. Discover Analytics, 4(1), Article 26. https://doi.org/10.1007/s44257-026-00075-4

Image Credits: AI Generated

DOI: 10.1007/s44257-026-00075-4

Keywords: supply chain optimization, machine learning, Random Forest, blockchain, multi-objective optimization, carbon emissions, linear programming, sustainability, predictive analytics, traceability, order fulfillment, enabled