ORIGINAL RESEARCH article

Front. Mater., 22 October 2026

Sec. Structural Materials

Volume 13 - 2026 | https://doi.org/10.3389/fmats.2026.1950801

EPD-grounded carbon-resilience allocation for low-carbon concrete supply: decision fidelity, route validation, and correlated-disruption stress testing

  • 1. Department of Civil Engineering, Chengdu Technological University, Chengdu, Sichuan, China

  • 2. College of Intelligent Construction and Manufacturing, Sichuan Technology and Business University, Chengdu, Sichuan, China

Abstract

Introduction:

Green construction teams need a defensible way to convert product-specific concrete carbon declarations into supply decisions that recognize delivery and disruption exposure. Prediction-to-optimization workflows can obscure a critical evidence boundary: verified environmental product declaration (EPD) values can support formal supplier comparison, whereas machine-learning estimates require decision-level validation before allocation.

Methods:

We introduce an evidence-aligned framework that separates observed EPD attributes, model-estimated screening outputs, scenario-derived A4 transport, and assumption-based disruption stress and assigns each a specific decision role. Model transportability was audited using plant, company, temporal, and geographic holdouts, followed by a direct comparison of predicted and EPD-observed rankings and allocations. Formal strength-constrained optimization minimized observed EPD A1–A3 plus scenario A4 under plant-share caps.

Results:

After harmonization and deduplication, the analysis contained 45,739 EPD records from 663 plants. A prospective external cohort added 22 declarations from 14 plants and seven derivation-unseen companies; the frozen model reduced record-level MAE by 6.6% relative to the frozen training median, while the supplier-ranking gate retained the EPD-grounded path. Across nine city-strength cases, a 40% cap produced three-plant allocations with a mean Herfindahl–Hirschman Index of 0.36. Under a 10% independent plant-disruption stress rate, it reduced the probability of losing at least half of supply from 0.100 to 0.027; mean and median A1–A4 premiums were 14.57% and 7.76%, respectively. The benefits remained measurable across disruption probabilities of 5%–30% and pairwise correlations of 0–0.50. An OpenStreetMap routing check covered 79 plant–project pairs: median distance error was 6.39%, 17 of 18 allocation decisions were unchanged, and the remaining case had 0.026% route-evaluated objective regret.

Discussion:

The contribution is a transparent carbon-concentration frontier with an empirical decision-fidelity gate: prediction supports early screening, whereas EPD values anchor formal allocation. The framework provides extensible planning evidence while preserving clear boundaries between comparative scenarios and operational procurement.

1 Introduction

The transition to nearly zero-emission buildings requires that project management address material-related greenhouse gas emissions alongside operational energy. Concrete is central to this transition because its production scale makes incremental improvements consequential across buildings and infrastructure (United Nations Environment Programme, 2024; Röck et al., 2020; Habert et al., 2020). Building-scale reviews further show that comparable system boundaries and transparent embodied-carbon inventories are prerequisites for meaningful material decisions (Cabeza et al., 2014; De Wolf et al., 2017). At the material scale, established pathways include clinker substitution, supplementary cementitious materials (SCMs), alternative binders, production efficiency, carbon capture, service-life extension, and specification-informed procurement (Scrivener et al., 2018; Miller et al., 2018; Gartner and Hirao, 2015). Current research is making these pathways more actionable through multi-criteria SCM screening and lithium-slag synthesis (Akintayo et al., 2025; Razzazan et al., 2025), experimental valorization of calcined fine coal gasification slag and ready-mix wash water (Lv and Fu, 2026; Zhou et al., 2026), and waste-derived, geopolymer, and lifecycle-aware concrete design (Li et al., 2026; Zhang et al., 2026; Tan et al., 2026b). Together, this literature establishes an expanding set of technically credible lower-carbon products. The project-management question is how to translate these product-level opportunities into a geographically and structurally eligible supply plan.

EPDs provide standardized, product-specific environmental information under life-cycle assessment and product-category rules (International Organization for Standardization, 2006a; International Organization for Standardization, 2006b; 2017). For ready-mixed concrete, procurement screening must reconcile declared units, validity dates, strength classes, plant locations, and multiple mixtures per plant. It must also distinguish product-stage A1–A3 emissions from delivery-stage A4 emissions. A low A1–A3 value does not automatically produce a low delivered-carbon option, yet an A4 value based on representative coordinates and generic truck assumptions is not equivalent to an operational shipment inventory. An effective framework therefore needs both a combined comparison and an explicit account of the evidence that supports each component.

Concrete machine learning developed from nonlinear strength modeling based on mixture proportions (Yeh, 1998) and now spans mechanical-property prediction, carbon estimation, and material sustainability assessment (Ben Chaabene et al., 2020; Wang et al., 2024). Recent explainable machine-learning studies combine prediction with lifecycle or sustainability interpretation across cementitious and asphalt materials (Xing et al., 2026; Tan et al., 2026a; b). This progression demonstrates the value of nonlinear models but leaves a distinct decision problem: predictive accuracy is not, of itself, evidence that predicted supplier rankings or allocations reproduce decisions based on reported EPD values. The distinction is central to “predict, then optimize” systems because downstream selection can amplify errors that are modest in aggregate metrics (Bertsimas and Kallus, 2020; Elmachtoub and Grigas, 2022).

Construction supply-chain studies likewise show that logistics, production, and project coordination must be considered jointly (Vrijhoef and Koskela, 2000; Naso et al., 2007). Recent low-carbon and disruption-aware studies add environmental and resilience objectives (Peng et al., 2025; Shi and Zhu, 2025; Hettiarachchi et al., 2025). However, public EPD records do not provide verified contract capacity, price, time-window, or disruption-frequency fields. This creates a scientific opportunity that is narrower and more useful than treating an open dataset as a complete procurement system: the data can support an evidence-grounded carbon-concentration analysis, provided that observed, estimated, and assumed quantities are not conflated.

This study addresses that opportunity through three linked contributions. First, it introduces a decision-fidelity gate that compares prediction-based supplier rankings and allocations directly with their EPD-observed counterparts under plant, company, temporal, geographic, and prospective producer-disjoint external validation. The gate assigns a prediction to the decision role supported by the evidence. Second, it formulates an EPD-grounded allocation model whose objective and reported carbon metric are identical: observed A1–A3 plus a clearly defined A4 scenario, subject to strength eligibility and a tunable maximum plant share. Third, it maps the resulting carbon-concentration frontier under both independent and correlated disruption stress and tests the transport approximation against network-routed distances. The framework is evaluated for Los Angeles, Atlanta, and Newark at 30, 40, and 50 MPa strength floors. Its main output is not a universal supplier ranking; rather, it is an auditable method for locating a project-specific balance between declared carbon and supply concentration.

2 Materials and methods

2.1 Evidence-aligned study design and system boundary

Figure 1 separates the formal allocation path from the prediction-audit branch. Table 1 defines the four evidence classes before model construction. Observed EPD A1–A3 and declared compressive strength determine formal eligibility and product-stage carbon. Cross-fitted A1–A3 predictions are evaluated as a pre-EPD or an incomplete-information screening aid. Scenario-derived A4 provides a common planning comparison. Plant-share limits and disruption parameters expose resilience trade-offs without being presented as empirical disruption forecasts.

FIGURE 1

TABLE 1

Evidence classVariablesRoleClaim boundary
Observed public EPDA1–A3 GWP, declared strength, plant and company, issue and validity datesFormal eligibility and carbon-allocation objectiveComparisons among represented, valid EPD products
Model-estimatedCross-fitted A1–A3 GWPEarly screening and decision-fidelity auditScreening support when a verified product value is unavailable
Scenario-derivedProject point, distance, round trip, truck factor, and payloadComparable A4 planning coefficientComparative scenario; not an operational delivery estimate
Assumption-based stressPlant-share cap, disruption probability, and correlationCarbon-concentration frontierConditional stress response; not an observed disruption-frequency claim

Evidence classes, analytical roles, and claim boundaries.

The carbon boundary comprises reported product-stage A1–A3 plus modeled delivery-stage A4, expressed as kg e . Construction pumping, site waste, use stage, carbonation, maintenance, end of life, and benefits beyond the system boundary are outside this comparison. The decision unit is a continuous allocation share of a 1,000 project demand. Three project representative points—Los Angeles, California, Atlanta, Georgia, and Newark, New Jersey—create contrasting candidate pools. Minimum declared compressive strengths are 30, 40, and 50 MPa. Detailed parameters, data-quality checks, validation results, decision-fidelity analyses, allocation outcomes, disruption sensitivity, and route checks are provided in Supplementary Material (Data Sheet 1, Tables S1-S10).

2.2 Public EPD data, harmonization, and deduplication

The frozen source table contained 47,413 publicly-linked ready-mixed concrete EPD rows, including company, plant, mixture label, declared strength, declared unit, A1–A3 GWP, issue and validity dates, program operator, application, and source link. The evidence cutoff of 23 July 2026 denotes the EPD-validity and source-freeze date; it is not a prediction horizon or project-delivery date. Records were retained when A1–A3 GWP was positive, strength was 5–150 MPa, and the declared unit was cubic meter or yard.

Strength in pounds per square inch was converted using Equation 1:and GWP per cubic yard was converted as . Data quality control removed 48 raw exact duplicate rows and, after harmonization, 212 additional rows identical across all analysis variables. The final analytical table contained 45,739 records, 663 plant groups, 83 companies, and nine US regions, with no missing target or strength values and no remaining analysis-equivalent duplicates. A plant group was defined by company and plant name. The second deduplication prevents records that differ in unused source fields from receiving repeated analytical weight.

The formal candidate layer used plants with primary US Census address matches and active EPD records. It contained 15,103 product rows and 81 plants before project and strength filtering. Candidate support was 24, 20, and 18 plants for Los Angeles, 21, 11, and 9 for Atlanta, and 34, 33, and 30 for Newark at the 30, 40, and 50 MPa floors, respectively.

2.3 Geography and A4 transport scenario

Representative street, city, state, and five-digit ZIP addresses were submitted to the US Census Geocoder (U.S. Census Bureau, 2026). Primary coordinates consisted of 386 exact and 135 non-exact address-range matches. Zip and city representative points were retained for geographical sensitivity records but did not enter formal allocation. Great-circle distance between plant and project was calculated using the Haversine formula. The planning road-distance approximation was defined by Equation 2:in miles. Products beyond 75 miles were excluded. Nominal A4 intensity was calculated using Equation 3:where “2” represents a round trip, is nominal payload, and kg e per vehicle-mile was assembled from the US EPA 2025 factor components (U.S. Environmental Protection Agency, 2025).

Equation 3 is a deliberately transparent planning estimate. Real delivery emissions depend on road choice, traffic, truck, and fuel technology, batching sequence, return-load practice, and payload. To test the decision sensitivity attributable specifically to distance, all 79 eligible plant–project pairs at the 30 MPa floor were queried through the Open Source Routing Machine table service using its driving profile (Project OSRM, 2026). The network-routed distances then replaced Equation 2, while the emission factor and payload remained fixed. Using the network-routed objective, we re-solved and evaluated 18 allocations—two share caps across nine project-strength cases.

2.4 Prediction audit under plant, company, temporal, and geographic shift

The prediction target was reported as A1–A3 GWP. Numeric features were declared compressive strength and EPD issue year. Categorical features were plant state, US region, EPD program operator, primary application, and where specified, company. The histogram gradient-boosting regressor used absolute-error loss, 250 iterations, learning rate 0.07, a maximum of 31 leaf nodes, minimum leaf size 30, and regularization of 1.0 (Friedman, 2001; Pedregosa et al., 2011). Missing numeric and categorical inputs were median- and most-frequent-imputed, respectively; an explicit code represented unseen categories.

Four validation designs tested distinct use cases. Five-fold plant-group holdout evaluated unseen plants while permitting other plants from a represented company. Five-fold company holdout prevented any company from crossing training and test sets and was run with and without company as an input. Temporal validation was trained on 2021–2022 records and tested on 2023–2025 records, without company. Geographic validation left one US region out at a time, also without company. Metrics were mean absolute error (MAE), plant-macro MAE, root mean squared error (RMSE), , and mean observed-minus-predicted bias.

2.5 Prospective producer-disjoint external validation

After the feature set, learner, hyperparameters, and decision gate were frozen, the live public NRMCA EPD registry was accessed on 15 August 2026 through its product lookup records. The derivation table’s latest issue date was 3 November 2025. Candidate records were required to have a registry issue timestamp after that date, a current product PDF, and no producer match to the derivation companies. Producer separation combined normalized exact-name screening with a prespecified normalized-name similarity audit (threshold 0.75) to exclude parent and alias variants. Current versions were then de-duplicated by company, plant, and mix code. Each PDF retained had to report a date later than the derivation maximum, a declared unit of 1 , compressive strength of 5–150 MPa, and A1–A3 GWP of 1–2,000 kg e . This yielded 22 product declarations from 14 plants and seven companies, dated 3 December 2025 to 5 August 2026.

The company-excluded predictor was fitted once to all derivation records and applied to the external cohort without model selection, recalibration, or refitting. A frozen derivation-set median provided a no-feature benchmark. Record-level and plant-macro MAE, RMSE, , and bias were calculated; 95% intervals used 2,000 company-cluster bootstrap resamples. The external decision audit repeated plant ranking and top- overlap at 20, 25, and 30 MPa floors. The registry operator is shared with part of the derivation data, but every external product, issue period, plant, and producer in this cohort is outside the derivation set; the design therefore tests prospective producer transfer under a common North American EPD framework rather than cross-PCR transfer.

2.6 Decision-fidelity gate

Aggregate prediction metrics were supplemented by supplier-level tests. For each of the nine city-strength cases, one candidate product per plant was selected under (i) cross-fitted predicted A1–A3 plus nominal A4 and (ii) observed EPD A1–A3 plus the same A4. Spearman correlation, Kendall correlation, and top- overlap for quantified ranking fidelity. The unconstrained and 40%-capped allocation problems were then solved with predicted and observed coefficients. Prediction-based allocations were evaluated using observed EPD A1–A4, yielding the decision regret defined in Equation 4:and allocation overlap . This gate defines the model’s decision role: formal allocation uses predicted coefficients only if their supplier-level fidelity was adequate for that purpose. Otherwise, verified EPD values retained priority and prediction served search-space screening.

2.7 EPD-grounded allocation and aligned objectives

For each project-strength case, the eligible product with the lowest observed A1–A4 coefficient at each plant represented that plant. Let be volume from plant , , its reported A1–A3 intensity, and the maximum plant share. The linear program was formulated as Equation 5, subject to the demand-balance constraint in Equation 6 and the plant-share bounds in Equation 7:The objective and reported carbon outcomes are therefore the same observed-EPD-plus-A4 quantity. The unconstrained EPD-minimum allocation uses ; the primary EPD-diversified allocation uses . Caps of 0.20, 0.30, and 0.50 provide frontier sensitivity. Supplier concentration is summarized by HHI, . The model expresses a carbon-concentration decision and does not impute unobserved commercial capacities or costs.

2.8 Independent and correlated disruption stress

The disruption analysis tested marginal plant probabilities and pairwise correlations using 20,000 reproducible stress realizations per allocation setting. For , plant disruptions were independent Bernoulli events. For , a beta Bernoulli construction generated exchangeable correlated events using Equations 8, 9:This construction preserves marginal probability and pairwise correlation . The shortfall share is , where . Outcomes were mean shortfall, 95th-percentile shortfall, probability of any shortfall, and probability . The last metric captures loss of at least half of planned supply and is the prespecified severe-shortfall measure.

3 Results

3.1 Generalization and decision fidelity define the prediction role

Table 2 and Figure 2 show that validation scope materially changes the error estimate. Plant holdout with company included achieved MAE of 52.38 kg e and . Removing company increased plant-holdout MAE to 56.75. Company holdout produced MAE 71.63 with company and 72.62 without it; temporal and geographic MAEs were 69.31 and 68.62, respectively. The cross-design pattern provides a direct use rule: plant-holdout performance characterizes screening among represented producer systems, whereas company, time, and region holdouts characterize transfer to new contexts.

TABLE 2

ValidationCompany featureMAEMacro MAERMSE
Plant holdoutIncluded45,73952.3852.8277.200.400
Plant holdoutExcluded45,73956.7555.3781.960.324
Company holdoutIncluded45,73971.6369.4298.930.015
Company holdoutExcluded45,73972.6267.0598.530.023
Temporal: 2021–2022 to 2023–2025Excluded10,51069.3160.8797.810.155
Leave-one-region-outExcluded45,73968.6274.0794.340.104

Out-of-sample A1–A3 prediction across validation designs.

FIGURE 2

Decision-level comparison clarified how that signal should be used (Table 3). Mean Spearman and Kendall correlations between predicted and observed plant rankings were 0.519 and 0.404. Top-10 overlap reached 0.689, supporting a shortlisting role, while top-1 overlap was 0.111. When predicted coefficients directly determined allocation, mean overlap with EPD-grounded allocations was 0.111 for the unconstrained policy and 0.200 for the 40% cap; observed-objective regret averaged 188.98 and 173.00 kg e , respectively. The decision-fidelity gate therefore assigns prediction to pre-EPD screening and retains observed EPD A1–A3 for formal supplier allocation.

TABLE 3

MetricResultDecision interpretation
Spearman rank correlation0.519Moderate ordering signal
Kendall rank correlation0.404Moderate pairwise ordering signal
Top-1/top-3/top-5/top-10 overlap0.111/0.259/0.422/0.689Useful expansion from exact choice to shortlist
Allocation overlap, /0.111/0.200EPD values retained for formal allocation
Mean observed-objective regret, /188.98/173.00kg e

Predicted versus observed decision fidelity across nine city–strength cases.

3.2 Prospective external validation confirms the decision gate

The external audit retained 22 complete product declarations from 14 plants and seven derivation-unseen companies (Table 4). The frozen predictor achieved record-level MAE 66.94 kg e (company-cluster bootstrap 95% interval 44.31–106.19), compared with 71.68 for the frozen derivation median—a 6.62% reduction. External plant-ranking results activated the prespecified role gate: at the 20 MPa floor, Spearman correlation was and top-three overlap was 0; at 25 MPa, the corresponding values were and 0. Consequently, no external model output was elevated to formal supplier ranking or allocation. The external cohort therefore supplies a prospective out-of-derivation test of the framework’s central safeguard: coarse predictive information can be audited without displacing a verified EPD coefficient when decision fidelity is not established.

TABLE 4

Audit itemResultInterpretation
External support22 products; 14 plants; 7 companiesProduct, plant, producer, and issue period disjoint from derivation data
PDF issue-date range3 Dec., 2025–5 Aug. 2026Later than the derivation maximum of 3 Nov., 2025
Frozen model MAE66.94 kg e 95% company-cluster bootstrap interval: 44.31–106.19
Frozen median MAE71.68 kg e Model record-level MAE was 6.62% lower
Supplier-ranking gateEPD anchor retainedExternal predictions did not enter formal optimization

Prospective producer-disjoint external validation and decision-role outcome.

3.3 EPD-grounded carbon-concentration frontier

All 45 allocation problems—nine cases across five share caps—were feasible and optimal. The unconstrained EPD-minimum selected one plant, whereas the 40% EPD-diversified allocation selected three plants in every case and reduced mean HHI from 1.00 to 0.36. Its mean A1–A4 intensity was 208.82 kg e compared with 187.66 for the unconstrained allocation. Relative premiums varied with candidate-pool structure; across the nine cases, the mean was 14.57% and the median was 7.76%.

Figure 3 makes the trade-off explicit. At and , the 40% cap reduced mean from 0.100 to 0.027. A 20% cap used five plants and reduced it to 0.009 with a mean carbon premium of 31.59%. The 50% point is shaped by the prespecified threshold: a two-plant 50–50 allocation records a severe event whenever either plant is disrupted, so it is not a concentration optimum for this loss definition. The 40% point is the first tested cap that requires at least three sources and moves the portfolio below the single-disruption severe-loss threshold.

FIGURE 3

3.4 Diversification benefits across disruption probabilities and correlations

Figure 4 reports the full disruption grid. For the 40% cap, severe-shortfall probability was 0.007, 0.027, 0.105, and 0.216 at independent disruption probabilities of 0.05, 0.10, 0.20, and 0.30. At , increasing from 0 to 0.25 and 0.50 changed the severe-shortfall probability from 0.027 to 0.067 and 0.088. The corresponding unconstrained values were approximately 0.100 across correlation settings because a one-plant allocation has no within-portfolio correlation mechanism. Thus, the 40% cap retains measurable severe-shortfall reduction throughout the tested correlation range, while the narrowing gap quantifies the exposure created by shared hazards.

FIGURE 4

The cap grid also provides an actionable extension path. At and , the 20%, 30%, and 40% caps produced severe-shortfall probabilities of 0.085, 0.087, and 0.088, respectively. Under highly shared disruptions, adding nominal suppliers has less leverage than under independent events; resilience planning should therefore combine concentration control with assessment of common upstream, regional, and infrastructure dependencies.

3.5 Network routing validates the A4 decision approximation

All 79 requested routes returned valid network distances. As shown in Figure 5 and Table 5, the circuity approximation averaged 29.80 miles versus 30.19 miles for network routing, with a mean bias of miles. Median absolute percentage error was 6.39%; the 90th and 95th percentiles were 19.48% and 27.39%, respectively. The largest percentage error occurred on a short route affected by the five-mile floor, which limits the practical meaning of relative error near zero distance.

FIGURE 5

TABLE 5

MeasureResult
Successful plant–project routes79/79
Mean approximate/routed distance29.80/30.19 miles
Mean distance bias miles
Median/P90/P95 absolute percentage error6.39%/19.48%/27.39%
Exact allocation matches17/18
Minimum allocation overlap0.80
Maximum network-evaluated objective regret0.057 kg e (0.026%)

Network-route validation and allocation sensitivity.

More importantly for the study objective, allocation decisions were stable. All nine unconstrained EPD-minimum allocations and eight of nine 40%-capped allocations were identical after network routing. The remaining Newark 50 MPa capped case retained 80% allocation overlap; evaluating the approximation-selected allocation with network distances produced regret of 0.057 kg e , or 0.026%. These results validate the approximation for the present comparative cases while maintaining the distinction between network routing and observed delivery operations.

Table 6 consolidates the empirical evidence supporting each decision layer. The role assignment is itself a study result: broad shortlist overlap supports efficient candidate discovery; reported EPD coefficients support formal allocation; route stability supports the comparative A4 approximation in the tested cases; the stress grid supports conditional resilience comparisons across explicit dependence assumptions.

TABLE 6

Decision layerPrincipal evidenceSupported use
Pre-EPD screeningMean top-10 overlap 0.689 across nine casesReduce a broad search space before product verification
Formal carbon allocationReported EPD A1–A3 plus common A4 scenarioRank eligible represented products and map share-cap premiums
A4 comparison17/18 route-recomputed decisions unchanged; maximum regret 0.026%Comparative planning within the tested city-strength cases
Resilience comparison5%–30% marginal probabilities and 0–0.50 correlationsEvaluate concentration policies conditional on explicit stresses

Evidence-to-decision summary for the framework.

4 Discussion

4.1 Scientific gap and distinctive capability

The central gap is not a lack of concrete carbon predictors or optimization algorithms in isolation; it is the absence of an evidence-aligned mechanism that decides which quantity is entitled to drive a supplier decision. The proposed decision-fidelity gate fills that gap. It tests model transportability beyond plant holdout, measures ranking and allocation fidelity against reported EPD values, and then assigns prediction to a role supported by those results. This prevents a convenient predicted coefficient from replacing a verified environmental declaration when both are available.

This role separation strengthens rather than discards the value of machine learning. Mean top-10 overlap of 0.689 shows that the model can narrow a broad pre-EPD search space using routinely available attributes. That capability is useful when a project team is exploring regions, strength classes, or provisional producer lists before collecting current product declarations. Once a valid EPD is available, the formal path switches to the observed value. The same gate is extensible: a future model with higher external ranking and allocation fidelity can earn a broader decision role without changing the framework.

The prospective producer-disjoint audit demonstrates that the gate operates beyond resampling within the compiled source table. Although the frozen predictor reduced record-level MAE relative to a frozen no-feature benchmark, the external ranking test retained the EPD anchor. This is the distinctive safety property of the framework: model transport can be tested at low cost, while allocation remains insulated from an unsupported expansion of model authority. The same audit can be rerun as new registry products appear, allowing the prediction role to expand only when external decision fidelity supports it.

The second distinctive capability is the aligned carbon-concentration frontier. The formal objective, the carbon result, and the route sensitivity all use the same EPD A1–A4 coefficient. Resilience enters through a transparent structural control—maximum plant share—and is evaluated across a prespecified loss measure. This makes the trade-off legible to project managers: the 40% cap creates a three-source portfolio and materially reduces the chance that one plant event removes at least half of planned supply, while its carbon premium remains directly observable for each project–strength case.

4.2 Implications for low-carbon concrete project management

The framework integrates three decisions that are often separated. First, strength eligibility precedes carbon ranking, so the comparison remains within the declared performance class. Second, A1–A3 and A4 remain separately traceable even though the allocation uses their sum. Project teams can therefore see whether an attractive option is driven by production carbon, transport, or both. Third, concentration is a selectable design variable rather than an after-the-fact statistic.

The route validation adds practical confidence at low analytical cost. Network routing changed one of 18 allocations, and the resulting objective difference was 0.026%. This does not make the A4 coefficient a field transport inventory; it demonstrates that the primary supplier decisions were not artifacts of the circuity approximation in the tested cases. A project moving toward quotation or contracting can replace the generic A4 component with carrier-specific route, truck, payload, and schedule data while retaining the same allocation structure.

Correlated stress also changes how diversification is interpreted. Independent plant events reward additional suppliers strongly; correlated events reduce that leverage because several suppliers can be affected together. The grid makes this dependence visible rather than embedding one disruption probability in a headline result. For construction management, the implication is constructive: supplier-count policies should be accompanied by a mapping of shared cement sources, common transport corridors, regional hazards, and other common-mode dependencies.

4.3 Decision-use boundary

The conclusions are supported within the defined evidence boundary. EPD A1–A3 values are reported environmental results rather than measurements of every delivered batch; their comparability depends on applicable product-category rules, declared units, background inventories, and validity. The dataset represents publicly available EPD products and is not a probability sample of the entire concrete market. The allocation conclusions therefore concern the represented candidate products and cases.

A4 is a scenario-based planning estimate. The routing check addresses network distance, but traffic, vehicle class, fuel, batching sequence, empty return, payload variability, and delivery timing require project-specific evidence. Accordingly, A4 results support comparative screening and methodological inference, not direct procurement, contracting, or shipment-accounting claims.

The plant-share cap is a resilience design parameter rather than verified production capacity. The disruption grid is conditional on stated probabilities and correlations; it estimates the behavior of allocation structures under common stress definitions and does not estimate a supplier’s real disruption frequency. Prices, capacities, lead times, time windows, quality consistency, contract terms, and minimum orders were not present in the public data. These fields can be added as constraints when a project owner obtains them.

Declared compressive strength is the formal performance filter. Durability, workability, early-age strength, shrinkage, modulus, exposure class, and service-life requirements should be added for project-specific structural specification. This modular boundary is an advantage for extension: additional verified criteria can enter eligibility without changing the evidence hierarchy or carbon–concentration logic.

5 Conclusion

This study establishes an evidence-aligned approach to low-carbon concrete supply planning. Its defining mechanism is a decision-fidelity gate that separates what a prediction model can shortlist from what verified EPD values should decide. Plant, company, temporal, geographic, prospective producer-disjoint, ranking, and allocation audits positioned machine learning as an efficient pre-EPD screening tool, while observed A1–A3 values anchored formal optimization. The 22-product prospective external cohort exercised this gate outside the derivation table: record-level MAE improved over a frozen no-feature benchmark, and the supplier-ranking audit retained the verified-EPD decision path.

The EPD-grounded allocation produced a transparent carbon-concentration frontier. Across nine city-strength cases, a 40% plant-share cap created three-source portfolios, reduced HHI from 1.00 to 0.36, and reduced severe shortfall probability from 0.100 to 0.027 under the primary independent 10% disruption stress. Correlated-disruption analysis showed where the benefit persists and where shared hazards become the more important control. Network routing further showed that the distance approximation preserved 17 of 18 allocation decisions and introduced at most 0.026% route-evaluated objective regret in the tested cases.

The resulting framework offers a clear capability for green construction management: it converts public environmental declarations into strength-constrained, location-aware, resilience-transparent planning evidence without overstating estimated quantities as observed operations. Its modular structure can incorporate exact routes, fleet data, verified capacity, cost, schedule, durability, and contract constraints as those data become available, providing a scalable path from public-data screening to project-specific low-carbon supply design.

Statements

Data availability statement

Publicly available datasets were analyzed in this study. These data can be found here: The publicly available ready-mixed concrete EPD records used in this study can be accessed through the ASTM International EPD directory (https://www.astm.org/standards-and-solutions/certification/environmental-product-declarations/epd-pcr) and the NRMCA EPD directory (https://www.nrmca.org/association-resources/sustainability/environmental-product-declarations/). Geographic data are available from the US Census Geocoder API (https://geocoding.geo.census.gov/geocoder/Geocoding_Services_API.html) and the 2025 Census Gazetteer files (https://www2.census.gov/geo/docs/maps-data/data/gazetteer/2025_Gazetteer/). The transport emission factors are available from the US EPA GHG Emission Factors Hub (https://www.epa.gov/climateleadership/ghg-emission-factors-hub). The trained models, analysis code, and derived outputs are available from the corresponding author upon reasonable request.

Author contributions

XY: Conceptualization, Data curation, Formal Analysis, Methodology, Writing – original draft. YX: Conceptualization, Project administration, Software, Supervision, Validation, Writing – original draft, Writing – review and editing. YW: Data curation, Formal analysis, Software, Writing – original draft.

Funding

The author(s) declared that financial support was not received for this work and/or its publication.

Acknowledgments

The authors gratefully acknowledge the organizations and institutions that maintain and provide public access to the datasets used in this study.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmats.2026.1950801/full#supplementary-material

SUPPLEMENTARY DATA SHEET 1

Supplementary methods and results, including study parameters, data-quality checks, internal and prospective external validation, supplier-ranking and allocation fidelity, carbon–concentration outcomes, disruption sensitivity, network-route validation, and reproducibility information.

References

  • 1

    AkintayoB. D.BabatundeO. M.AkintayoD. C.OlanrewajuO. A. (2025). Transforming industrial waste into low-carbon cement: a multi-criteria assessment of supplementary cementitious materials for sustainable concrete design. Recycling10, 211. 10.3390/recycling10060211

  • 2

    Ben ChaabeneW.FlahM.NehdiM. L. (2020). Machine learning prediction of mechanical properties of concrete: critical review. Constr. Build. Mater.260, 119889. 10.1016/j.conbuildmat.2020.119889

  • 3

    BertsimasD.KallusN. (2020). From predictive to prescriptive analytics. Manag. Sci.66, 1025–1044. 10.1287/mnsc.2018.3253

  • 4

    CabezaL. F.RincónL.VilariñoV.PérezG.CastellA. (2014). Life cycle assessment (LCA) and life cycle energy analysis (LCEA) of buildings and the building sector: a review. Renew. Sustain. Energy Rev.29, 394–416. 10.1016/j.rser.2013.08.037

  • 5

    De WolfC.PomponiF.MoncasterA. (2017). Measuring embodied carbon dioxide equivalent of buildings: a review and critique of current industry practice. Energy Build.140, 68–80. 10.1016/j.enbuild.2017.01.075

  • 6

    ElmachtoubA. N.GrigasP. (2022). Smart “predict, then optimize”. Manag. Sci.68, 9–26. 10.1287/mnsc.2020.3922

  • 7

    FriedmanJ. H. (2001). Greedy function approximation: a gradient boosting machine. Ann. Statistics29, 1189–1232. 10.1214/aos/1013203451

  • 8

    GartnerE.HiraoH. (2015). A review of alternative approaches to the reduction of CO2 emissions associated with the manufacture of the binder phase in concrete. Cem. Concr. Res.78, 126–142. 10.1016/j.cemconres.2015.04.012

  • 9

    HabertG.MillerS. A.JohnV. M.ProvisJ. L.FavierA.HorvathA.et al (2020). Environmental impacts and decarbonization strategies in the cement and concrete industries. Nat. Rev. Earth and Environ.1, 559–573. 10.1038/s43017-020-0093-3

  • 10

    HettiarachchiI.RotimiJ. O. B.ShahzadW. M.KahandawaR. (2025). Bridging sustainability and performance: conceptualizing net-zero integration in construction supply chain evaluations. Sustainability17, 5814. 10.3390/su17135814

  • 11

    International Organization for Standardization (2006a). ISO 14040:2006 Environmental Management—life Cycle assessment—principles and Framework.

  • 12

    International Organization for Standardization (2006b). ISO 14044:2006 Environmental management—life Cycle assessment—requirements and Guidelines.

  • 13

    International Organization for Standardization (2017). ISO 21930:2017 Sustainability in Buildings and Civil Engineering works—core Rules for Environmental Product Declarations of Construction Products and Services.

  • 14

    LiH.GhahsarehF. M.GuoP.TanX.TengL.MengW.et al (2026). Valorization of solid wastes in concrete production: an overview of materials, properties, mechanisms, and emerging technologies toward upcycling. Carbon Footprints5, 16. 10.20517/cf.2025.115

  • 15

    LvS.FuB. (2026). Using calcined fine coal gasification slag as a novel supplementary cementitious material for portland cement. Front. Mater.13, 1778301. 10.3389/fmats.2026.1778301

  • 16

    MillerS. A.JohnV. M.PaccaS. A.HorvathA. (2018). Carbon dioxide reduction potential in the global cement industry by 2050. Cem. Concr. Res.114, 115–124. 10.1016/j.cemconres.2017.08.026

  • 17

    NasoD.SuricoB.TurchianoB.KaymakU. (2007). Genetic algorithms for supply-chain scheduling: a case study in the distribution of ready-mixed concrete. Eur. J. Operational Res.177, 2069–2099. 10.1016/j.ejor.2005.12.019

  • 18

    PedregosaF.VaroquauxG.GramfortA.MichelV.ThirionB.GriselO.et al (2011). Scikit-learn: machine learning in python. J. Mach. Learn. Res.12, 2825–2830.

  • 19

    PengJ.SuZ.LiuX.MaC. (2025). Promoting low-carbonization in the construction supply chain: key influencing factors and sustainable practices. Sustainability17, 3375. 10.3390/su17083375

  • 20

    Project OSRM (2026). OSRM API Documentation: table service. Available online at: https://project-osrm.org/docs/v5.23.0/api/(Accessed August 12, 2026).

  • 21

    RazzazanS.MashaanN. S.ParaskevaT. (2025). Lithium slag as a supplementary cementitious material for sustainable concrete: a review. Materials18, 3641. 10.3390/ma18153641

  • 22

    RöckM.SaadeM. R. M.BalouktsiM.RasmussenF. N.BirgisdottirH.FrischknechtR.et al (2020). Embodied GHG emissions of buildings—the hidden challenge for effective climate change mitigation. Appl. Energy258, 114107. 10.1016/j.apenergy.2019.114107

  • 23

    ScrivenerK. L.JohnV. M.GartnerE. M. (2018). Eco-efficient cements: potential economically viable solutions for a low-CO2 cement-based materials industry. Cem. Concr. Res.114, 2–26. 10.1016/j.cemconres.2018.03.015

  • 24

    ShiM.ZhuJ. (2025). Procurement optimization for manufacturing enterprises considering supply chain disruption risks and carbon emissions. Sustainability17, 3532. 10.3390/su17083532

  • 25

    TanX.XingJ.MahjoubiS.GuoP.WeiZ.WangY.et al (2026a). Explainable machine learning and life cycle assessment for sustainable design of fiber-reinforced asphalt concrete. J. Clean. Prod.547, 147759. 10.1016/j.jclepro.2026.147759

  • 26

    TanX.XingJ.WangY.QiuH.MahjoubiS.GuoP. (2026b). Explainable machine learning for predicting compressive strength of rubberized concrete: SHAP interpretation, lifecycle assessment, and design recommendations. J. Clean. Prod.538, 147338. 10.1016/j.jclepro.2025.147338

  • 27

    United Nations Environment Programme (2024). 2023 Global Status Report for Buildings and Construction: Beyond foundations—mainstreaming Sustainable Solutions to Cut Emissions from the Buildings Sector. 10.59117/20.500.11822/45095

  • 28

    U.S. Census Bureau (2026). Census geocoding services web interface and Api. Available online at: https://www.census.gov/data/developers/data-sets/Geocoding-services.html (Accessed July 23, 2026).

  • 29

    U.S. Environmental Protection Agency (2025). Ghg emission factors hub. Available online at: https://www.epa.gov/climateleadership/ghg-emission-factors-hub (Accessed July 23, 2026).

  • 30

    VrijhoefR.KoskelaL. (2000). The four roles of supply chain management in construction. Eur. J. Purch. and Supply Manag.6, 169–178. 10.1016/S0969-7012(00)00013-7

  • 31

    WangM.DuM.JiaY.ChangC.ZhouS. (2024). Carbon emission optimization of ultra-high-performance concrete using machine learning methods. Materials17, 1670. 10.3390/ma17071670

  • 32

    XingJ.TanX.LiY.JinD.GuoP.WangY.et al (2026). Interpretable machine learning for predicting splitting strength of asphalt concrete: insights from SHAP analysis. Materials19, 1636. 10.3390/ma19081636

  • 33

    YehI.-C. (1998). Modeling of strength of high-performance concrete using artificial neural networks. Cem. Concr. Res.28, 1797–1808. 10.1016/S0008-8846(98)00165-3

  • 34

    ZhangM.GuoP.TanX.DuJ.MengW.BaoY. (2026). Cradle-to-gate assessment and optimization of sustainable geopolymer concrete. J. Clean. Prod.538, 147387. 10.1016/j.jclepro.2025.147387

  • 35

    ZhouH.MaX.WangX.ZhangF. j.YuL. (2026). The effect of treatment way of wash water form ready-mixed concrete plant on the properties of mortar. Front. Mater.13, 1836853. 10.3389/fmats.2026.1836853

Summary

Keywords

construction supply chain, decision fidelity, environmental product declaration, green construction management, low-carbon concrete, robust allocation, supplier diversification, transport emissions

Citation

Yang X, Xiang Y and Wei Y (2026) EPD-grounded carbon-resilience allocation for low-carbon concrete supply: decision fidelity, route validation, and correlated-disruption stress testing. Front. Mater. 13:1950801. doi: 10.3389/fmats.2026.1950801

Received

28 July 2026

Revised

16 August 2026

Accepted

01 September 2026

Published

22 October 2026

Volume

13 - 2026

Edited by

Antonio Caggiano, University of Genoa, Italy

Reviewed by

Xiao Tan, Hohai University, China

Carlos Humberto Martins, State University of Maringá, Brazil

Updates

Copyright

*Correspondence: Yong Xiang,

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

Outline

Figures

Cite article

Copy to clipboard


Export citation file


Share article

Article metrics