Optimized supply chain routing cuts benchmark costs by less than one percent but exposes fragility

A new study published in Discover Analytics on 17 September 2026 reports that a carefully audited mixed-integer linear programming model can shave less than one percent off the reconstructed historical cost of routing a public logistics benchmark, while revealing that the optimized network depends heavily on a single carrier, a single origin port, and a single plant. The paper, authored by Md Raisul Islam Khan, Muhtasib Sarker Tahsin, and Abdullah Al Mamun, is deliberately framed not as a novel algorithm but as a reproducible decision-analytics workflow that makes its data coverage, assumptions, and computational behavior fully inspectable.

The researchers worked with the public Supply Chain Logistics Problem Dataset, released under a CC BY 4.0 license through Brunel University London Figshare. The workbook contains seven interconnected tables covering products per plant, plant-port links, customer restrictions, warehouse capacities, freight rates, and a historical order list. Before any analysis, the executed workflow verified sheet names, row counts, column counts, key data types, and a dataset checksum. Warehouse capacity was interpreted as documented daily order-processing capacity, consistent with the workbook documentation, and the historical order list was treated as a batch rather than a dated queue for a single operating day.

A central feature of the study is its baseline coverage audit. Of 9,215 historical orders, only 7,845, or 85.13 percent, had reconstructable freight costs and entered the savings comparison. The remaining 1,370 excluded records were not randomly scattered: all of them used carrier V444_1, DTD service, the PORT04-to-PORT09 lane, and shipment weights falling inside a gap between the 2.01–2.50 and 70.51–99.99 freight-rate bands. The authors emphasize that the reported saving applies only to the matched subset and cannot be generalized to the excluded orders, whose exclusion pattern could bias the estimate if those orders have systematically different routing opportunities.

After filtering by product eligibility, plant-port connections, carrier-service compatibility, customer restrictions, and weight-band coverage, and after pruning dominated duplicate options, the workflow generated 23,891 non-dominated real route options covering all 7,845 matched orders. A diagnostic dummy option was added for each order so that infeasibility would be measured explicitly rather than hidden by solver failure. The full model contained 23,891 real-route binary variables, 7,845 diagnostic variables, and 7,864 constraints, and was solved with the HiGHS solver through scipy.optimize.milp under a 300-second limit with zero relative MIP gap. The instrumented full solve reached the optimum in 1.121 seconds.

Because the order batch exceeded any single day of documented plant capacity, the researchers ran a complete sweep of an aggregate capacity multiplier H from 1 to 10, seeking the smallest value that eliminated diagnostic dummy assignments. A one-day interpretation left 5,492 orders unsupported, and unsupported orders declined monotonically until H = 7 became the first zero-dummy multiplier. The authors are careful to stress that H = 7 is an empirical cumulative-capacity threshold, not a verified day-by-day schedule: the model contains no promised-delivery penalties, order-release dates, backlog states, or service-level agreements, and an illustrative plant-day binning diagnostic assumed all orders were available at the start of the horizon.

At H = 7, the main service-preserving model routed all 7,845 matched orders without dummy assignments. Optimized cost was 14,842,335.53 against a reconstructed historical baseline of 14,965,998.85, a reduction of 123,663.32, or 0.8263 percent. Two executable service-policy ablations showed that most of the attainable improvement was already captured by the main model: allowing no-slower substitutions added roughly 0.375 percentage points, and unrestricted DTD/DTP substitution contributed only a further 0.014 points, bringing total savings to about 1.20 to 1.22 percent. The authors attribute the modest gain partly to warehouse cost dominating many order totals and to historical assignments that appear relatively efficient within the service-preserving feasible set.

The stress testing produced the study’s most striking findings. Uniform freight-rate changes of plus or minus 10 percent preserved full feasibility and moved optimized cost by only about 7,135 in either direction, and all multipliers from 0.80 to 1.20 kept every route feasible while cost changed approximately linearly. Capacity loss behaved very differently: a 15 percent uniform reduction left 248 orders unsupported, and at a 30 percent reduction the unsupported count rose to 1,324. Removing the dominant optimized carrier V444_0 left 5,451 orders, or 69.48 percent of the matched set, without valid routes, while removing the dominant origin port PORT04 left 6,610 orders, or 84.26 percent, unsupported. These are feasibility-loss results, not cost results, and they reflect structural dependence rather than price exposure.

Concentration diagnostics confirmed how fragile the optimized plan is. PLANT03 processed 6,501 orders, or 82.87 percent; carrier V444_0 served 5,588 orders, or 71.23 percent; and PORT04 originated 6,871 orders, or 87.58 percent, with corresponding Herfindahl-Hirschman indices of 0.691, 0.533, and 0.771. The median order had only two feasible real options, and large majorities of orders had exactly one feasible plant, carrier, or origin port after all source-table restrictions were applied. Exhaustive removal of every observed carrier and port showed uneven exposure: removing V444_1, V444_4, V444_5, or V444_8 created no dummy assignments, removing CUSTOMER_ARRANGED made exactly 854 customer-arranged-freight orders unsupported, and removing PORT09 left 173 orders unsupported.

A Monte Carlo experiment with 100 reproducible draws, using seed 20260731 and assumed lognormal freight multipliers with coefficients of variation of 0.05 globally and 0.10 by lane, added distributional context. Mean optimized cost was 14,842,228.80 with a standard deviation of 6,453.49, the simulated 95 percent confidence interval for the mean ran from 14,840,963.92 to 14,843,493.69, and all draws remained feasible, with cost exceeding the deterministic base in 51 percent of draws. The authors caution that these distributions are analytical assumptions used to probe cost dispersion, not empirical estimates of future freight volatility, and that the analysis does not estimate the probability of carrier failure, port closure, or capacity loss.

The study also declines to rank its objective against the best-known ant-colony-optimization result of 2,701,367.58 reported for the same benchmark by Dzalbs and Kalganova, noting that the present analysis uses a matched historical subset, order-by-order freight reconstruction, explicit customer-arranged-freight treatment, and an aggregate capacity multiplier, so a direct comparison would be invalid without harmonizing order scope, cost accounting, and capacity treatment. Among the stated limitations are the single public benchmark, the non-random exclusion of 1,370 orders, the aggregate rather than time-indexed capacity treatment, the absence of vehicle-tour optimization, and illustrative rather than empirically estimated shock magnitudes. The authors argue that the concentration findings are more consequential than the 0.83 percent saving: a low-cost plan that assigns most orders to one carrier and one origin port creates substantial continuity exposure, and the identified PORT04-to-PORT09 freight-band gap is a master-data issue that should be resolved before the model informs enterprise decisions.

Subject of Research: Supply chain logistics optimization using mixed integer programming and scenario simulation on a public benchmark dataset

Article Title: Supply chain logistics optimization using mixed integer programming and scenario simulation

Article References: Khan, M. R. I., Tahsin, M. S., & Mamun, A. A. (2026). Supply chain logistics optimization using mixed integer programming and scenario simulation. Discover Analytics, 4(1), Article 40. https://doi.org/10.1007/s44257-026-00091-4

Image Credits: AI Generated

DOI: 10.1007/s44257-026-00091-4

Keywords: supply chain, mixed integer programming, logistics optimization, Monte Carlo simulation, route concentration, capacity constraints, benchmark dataset, stress testing, freight costs, decision analytics, feasibility, HiGHS solver