
The blood supply chain is complex due to its perishable nature, time criticality, dependence on voluntary donations, and vulnerability to demand fluctuations, particularly during crises such as natural disasters and pandemics. This study proposes a decision making framework for sustainable and resilient BSC network design by combining fuzzy demand forecasting with multi objective robust optimization. The framework operates in two stages. First, a Long Short Term Memory (LSTM) neural network forecasts blood demand across five transfusion centers in Fars province, Iran, generating triangular fuzzy numbers for optimistic, baseline, and pessimistic scenarios. The forecasting model achieves a Mean Absolute Percentage Error (MAPE) of 4.42% across the five transfusion centers. Second, these fuzzy outputs are incorporated into a multi-objective, fuzzy-informed optimization framework that captures demand variability through scenarios and adopts a conservative upper-bound treatment of temporary-center disruption rates and environmental emission coefficients. The model integrates decisions on permanent and temporary center locations, donor allocation, and blood flow management while minimizing total cost, blood shortage, and environmental emissions. The model is solved using the Non dominated Sorting Genetic Algorithm II (NSGA II), producing a converged (GD = 0.07) and uniformly distributed (Spread = 0.08) Pareto front with seven non dominated solutions. Validation against CPLEX and MOPSO on small scale instances confirms the reliability of NSGA II, with an optimality gap below 0.5%. Sensitivity and resilience analyses show that the designed network maintains a 99.3% demand fulfillment rate under severe disruption scenarios while allowing temporary center locations and resource allocations to adapt to changing conditions. The framework provides a data driven decision support tool for healthcare policymakers and offers a transferable approach for resilient healthcare network planning. By preserving uncertainty throughout the forecasting and optimization pipeline, the study strengthens the connection between data driven prediction and real world network design.
Circular supply chain network design requires not only the coordination of forward and reverse networks but also an explicit representation of how products are manufactured, recovered, and reintegrated across multiple life cycles. However, many existing models often treat manufacturing and recycling in an aggregated manner, limiting their ability to capture operational circularity, particularly under uncertainty. This study proposes a multi-period, multi-product network design model for circular supply chains that integrates forward and reverse bill of materials with manufacturing and recycling process networks. The model jointly optimizes facility locations, technology choices, capacity levels, and commodity flows. Uncertainties in product demand, material supply, return rates, returned product quality, and economic parameters are represented using fuzzy numbers to capture epistemic uncertainty when reliable probability distributions are difficult to establish. A modified hybrid robust possibilistic chance-constrained programming approach is developed using possibility and credibility measures, together with a linearization scheme to ensure consistency in feasibility degrees. The applicability of the model is demonstrated through a lead-acid battery supply chain case study, followed by comparative and sensitivity analyses. Under deterministic conditions, the integrated circular network obtains a 7.68% reduction in total cost compared with the linear network. Furthermore, an evaluation across different uncertainty realizations shows that the proposed robust fuzzy approach obtains lower average total cost and standard deviations than its conventional fuzzy and deterministic counterparts. These results demonstrate the potential of the proposed framework in improving cost efficiency and solution stability while supporting strategic circular supply chain decisions under uncertainty.
Stakeholder-perception surveys are central to meso-economic analysis of technology diffusion in agri-food supply chains, but long multi-domain questionnaires impose respondent burden, particularly along Geographical-Indication (GI) and organic value chains. This paper proposes a Tabu Search framework that treats short-form item selection as a combinatorial optimisation problem, using the segmentation structure of the full instrument as the reference to preserve. Candidate subsets are scored by a composite fidelity objective combining (i) Procrustes similarity between the principal-component scores of the reduced and full instruments and (ii) the Adjusted Rand Index (ARI) of the induced stakeholder-profile assignments, with a sparsity penalty. The method is applied to a 40-item survey of 42 stakeholders from seven GI and organic pilot value chains in seven European countries, eliciting perceptions of traceability and anti-fraud innovations. Principal Component Analysis retains four components (68.1% of variance), and Ward hierarchical clustering identifies four profiles: Pragmatic Optimists, Committed Enthusiasts, Operationally Cautious, and Risk-Averse Sceptics. Across 30 runs the algorithm converges to short forms of seven to eleven items that recover the segmentation with perfect agreement (ARI = 1.000) and high geometric fidelity (mean Procrustes 0.907). The best-of-run solution retains seven items and attains a higher composite objective than eight alternative wrapper, filter, and embedded procedures evaluated at the same size. Repeated subsampling (200 subsamples of the same records) shows that perfect profile recovery is reproduced in every subsample, with Q8 and Q21 forming a robust core. The contribution is twofold: methodological, providing a metaheuristic item-selection framework portable across stakeholder surveys in supply chain management; and practical, delivering an internally validated candidate short form for monitoring adoption readiness for traceability innovations in agri-food chains.
This study develops a capacity-constrained Stackelberg framework for a three-channel renewable microgrid supply chain consisting of one upstream microgrid, two geographically separated distribution-network-operator franchise channels, and a distinct direct-retail segment. The model jointly considers wholesale pricing, downstream retail markups, service quality, local advertising, and a shared renewable-energy feasibility limit. Physical feasibility is incorporated directly into both decentralized and centralized optimization through Karush-Kuhn-Tucker conditions, while an active-set formulation characterizes channel shutdown under severe scarcity. The analytical results show that vertical integration eliminates the double-marginalization effect in the intermediated channels and, when renewable capacity is slack, produces an exact doubling relationship in downstream demand, service quality, and advertising effort without requiring symmetry between the two DNOs. When the shared renewable-energy constraint binds, a positive scarcity shadow price limits demand expansion, weakens the integration advantage, and can eventually render individual market channels inactive. For interior capacity-slack equilibria, higher marginal generation and delivery cost reduces demand and operating profit, whereas greater service-quality effectiveness improves channel performance and higher quality-provision cost weakens it. Under binding capacity with an unchanged active set, a marginal change in generation cost may instead be offset by the scarcity multiplier. Renewable availability affects outcomes primarily through feasibility thresholds and regime transitions once capacity becomes scarce. The numerical analysis is illustrative rather than empirical and is used to validate the analytical structure, threshold behavior, and managerial implications of the model.
Traditional machine learning (ML) techniques are widely used to support operational supply chain decisions, but their “black box” nature limits transparency and actionability. This study addresses this challenge by exploring the effectiveness of causal discovery methods and large language models (LLMs) in supply chain management with a focus on the automotive industry. Six causal discovery algorithms- Peter–Clark (PC), Fast Causal Inference (FCI), Greedy Equivalence Search (GES), DirectLiNGAM, NOTEARS, and Regression with Subsequent Independence Test (RESIT) are benchmarked against an expert-validated ground-truth graph constructed from 18 domain experts with a combined 300+ years of experience, using 12 months of internal supplier Key Performance Indicator (KPI) data. RESIT achieves the closest alignment with expert knowledge, with a False Discovery Rate of 0.1429, Structural Hamming Distance of 1, precision and recall of 0.8571 and an F1-score of 0.85, outperforming all other algorithms on every metric. Robustness of the estimated causal effect of Backlog on Escalation is confirmed through three refutation tests-random common cause, random subset and placebo treatment returning high p-values of 0.96, 0.76, and 0.98 respectively, indicating insensitivity to spurious confounders and random perturbations. Counterfactual simulation shows that reducing Backlog by 50%–70% consistently lowers escalation levels across suppliers. LLMs contribute by extracting causal factors from unstructured external event data, accelerating causal graph construction, and reducing dependence on manual expert input. LLM-derived causal graphs are further compared with expert-validated models to ensure alignment and reliability, and are operationalized through a Weighted Sum Model risk score linking external drivers to internal KPIs for actionable supplier escalation decisions.