Feedback-enabled digital twin and explainable artificial intelligence framework for predictive resilience and adaptive decision-making in disrupted supply chains

1 Introduction

Supply chains have become increasingly exposed to frequent, complex, and interrelated disruptions, particularly in Nigeria and other developing economies where operational systems often function under unstable infrastructural, economic, environmental, and institutional conditions (Ivanov and Dolgui, 2021; Wang et al., 2022). In such environments, firms are required to move raw materials, finished goods, information, and financial resources across networks that are frequently affected by fuel scarcity, port congestion, insecurity, poor road conditions, exchange-rate volatility, supplier unreliability, weather-related delays, policy changes, demand instability, and sudden logistics breakdown (Ivanov and Dolgui, 2021; Bhandal et al., 2022; Ivanov, 2020). These disruptions do not only delay delivery; they weaken inventory availability, increase procurement and transportation costs, reduce service reliability, and expose firms to customer dissatisfaction, revenue loss, and operational uncertainty. For many Nigerian industries, especially in manufacturing, agro-processing, pharmaceuticals, retail, and food distribution, supply-chain disruption has therefore become a major threat to competitiveness and business continuity (Wang et al., 2022; Queiroz et al., 2022).

Traditional supply-chain resilience practices are useful but increasingly inadequate. Most existing approaches emphasize monitoring, response, and recovery after disruption has already occurred (Ivanov, 2020). Managers observe stock-outs, late deliveries, rising lead times, route failures, supplier defaults, and cost overruns, and then attempt to respond through emergency procurement, supplier switching, shipment rescheduling, or manual inventory redistribution (Bhandal et al., 2022; Nwamekwe and Igbokwe, 2024). This approach is reactive and often costly because decisions are made when operational damage has already started. In practical terms, a firm may only discover its vulnerability after warehouse stock has dropped below safety level, a supplier has failed to deliver, or a critical transport route has become unavailable. Such delayed response cannot provide the level of agility required in a modern supply-chain environment where disruption signals emerge gradually and interact across several nodes before becoming visible as performance failure (Hosseini et al., 2019; Belhadi et al., 2022).

The growing availability of digital technologies provides a stronger pathway for transforming supply-chain resilience from a reactive function into a predictive and adaptive capability (Hosseini et al., 2019). Digital twins, machine learning, explainable artificial intelligence, real-time data analytics, uncertainty quantification, and optimization models can jointly support a new generation of intelligent supply-chain systems. A digital twin can create a virtual representation of the physical supply chain by modelling suppliers, production facilities, warehouses, transport routes, customer locations, inventory nodes, lead-time behaviour, demand flow, cost structure, disruption events, and service-level performance. When properly connected to real-time or near-real-time data streams, the twin becomes a living model that can mirror current operations, simulate future disruption scenarios, and test alternative recovery strategies before implementation (Queiroz et al., 2022; Bhandal et al., 2022; Belhadi et al., 2022).

Machine learning strengthens this digital twin architecture by learning hidden patterns from historical and live supply-chain data. Models such as Random Forest, XGBoost, and Long Short-Term Memory networks can predict disruption vulnerability, stock-out risk, supplier delay, demand surge, delivery failure, and service-level deterioration. However, prediction alone is not enough (Baryannis et al., 2019). A model may indicate that disruption risk is high, but the decision-maker still needs to know why the risk is high, how reliable the prediction is, and which recovery option will produce the best operational outcome. This is where explainable artificial intelligence and uncertainty quantification become critical. SHAP-based explainability can show whether a predicted disruption is mainly driven by supplier unreliability, low inventory cover, port delay, route instability, demand fluctuation, transport-cost escalation, or production-capacity limitation. Uncertainty quantification can further indicate whether the prediction is highly reliable or whether the system should recommend cautious monitoring due to weak confidence (Hosseini et al., 2019; Olan et al., 2024).

This study therefore proposes a feedback-enabled digital twin and explainable artificial intelligence framework for predictive resilience and adaptive decision-making in disrupted supply chains. The framework is designed to move supply-chain management from the conventional cycle of monitor, respond, and recover to a more intelligent cycle of sense, predict, simulate, adapt, and learn (Baryannis et al., 2019; Hosseini et al., 2019). The central novelty is the development of a closed-loop supply-chain intelligence system that integrates real-time supply-chain data streams, disruption early-warning detection, machine learning prediction, SHAP explainability, uncertainty-aware risk evaluation, digital twin scenario simulation, multi-objective recovery optimization, and feedback-based learning. This closed-loop logic ensures that the system does not merely detect supply-chain problems, but continuously learns from operational outcomes and improves future decisions (Nwamekwe and Igbokwe, 2024; Busse et al., 2021).

The proposed framework represents the physical supply chain as a dynamic and traceable digital model. Operational data from enterprise resource planning systems, warehouse records, transport logs, supplier reports, procurement histories, demand records, weather information, port-clearance reports, and disruption incident logs are processed through a data layer (Belhadi et al., 2022; Queiroz et al., 2022). The artificial intelligence prediction engine then estimates disruption vulnerability and performance degradation. The digital twin receives these outputs and simulates alternative recovery actions under defined operational constraints. The risk evaluation layer compares expected delay, stock-out probability, service-level loss, recovery cost, and confidence range (Ivanov, 2020). The adaptive decision layer then recommends feasible recovery strategies, including supplier switching, order reallocation, route diversion, emergency procurement, inventory redistribution, production rescheduling, safety-stock adjustment, and transport-mode substitution. After implementation, actual outcomes are compared with predicted and simulated outcomes, and the feedback is used to update the learning system (Olan et al., 2024; Binsfeld and Gerlach, 2022).

The study is guided by five objectives. The first is to develop a digital twin representation of supply-chain operations. The second is to predict disruption vulnerability using machine learning models. The third is to quantify uncertainty in supply-chain decisions. The fourth is to identify dominant disruption drivers using explainable artificial intelligence. The fifth is to develop adaptive recovery and resilience strategies for disrupted supply chains. These objectives are structured to ensure that the framework is not only computationally intelligent but also operationally useful and managerially interpretable.

The contribution of the study is both theoretical and practical. Theoretically, it advances supply-chain resilience research by integrating prediction, simulation, explainability, uncertainty, optimization, and learning into one closed-loop framework (Belhadi et al., 2022; Maheshwari and Kamble, 2022). Practically, it provides a decision-support pathway for firms operating in uncertain environments such as Nigeria, where supply-chain disruptions are regular and often difficult to control manually (Kritzinger et al., 2018). By enabling firms to sense weak signals, predict vulnerability, simulate recovery options, explain risk drivers, quantify decision confidence, and adapt based on feedback, the framework offers a strong foundation for building cognitive resilient supply chains that are proactive, transparent, and self-improving (Hosseini et al., 2019).

This study differs from existing digital twin and AI-based supply-chain resilience studies by proposing a closed-loop cognitive resilience architecture that integrates six functions in one operational cycle: disruption sensing, machine-learning prediction, explainable driver diagnosis, uncertainty-aware risk evaluation, digital twin scenario simulation, and feedback-based model updating. Unlike studies that apply digital twins mainly for representation or simulation, the present framework treats the digital twin as an adaptive decision engine. Its methodological contribution lies in linking predictive analytics, SHAP-based explanation, Monte Carlo uncertainty assessment, multi-objective recovery optimization, and feedback recalibration into a single sense–predict–simulate–adapt–learn system. This makes the framework conceptually different from conventional monitor–respond–recover resilience models.

1.1 Critical synthesis of existing studies and research gap

Existing supply-chain resilience studies have made important contributions in digital twin modelling, disruption simulation, artificial intelligence prediction, and risk management. However, most studies still treat these components separately. Digital twin studies often focus on virtual representation and scenario testing, while machine-learning studies emphasize predictive accuracy without sufficient explainability, uncertainty assessment, or adaptive recovery validation. Similarly, resilience studies commonly retain a reactive monitor–respond–recover logic, where decisions are made after disruption signals have already affected operational performance. The major research gap, therefore, is the absence of an integrated, explainable, uncertainty-aware, and feedback-enabled digital twin framework that can predict disruption vulnerability, explain its causes, simulate recovery options, optimize decisions, and update itself through feedback. The present study addresses this gap by developing a closed-loop cognitive framework for disrupted supply chains.

2 Materials and methods

2.1 Research design

This study adopted a quantitative modelling, simulation, and decision-optimization design to develop a feedback-enabled digital twin and explainable artificial intelligence framework for predictive resilience and adaptive decision-making in disrupted supply chains. The design was structured to achieve five linked methodological tasks: digital representation of supply-chain operations, disruption-vulnerability prediction, uncertainty quantification, explainable risk diagnosis, and adaptive recovery recommendation (Tao et al., 2018). A closed-loop logic was applied so that operational data from the physical supply chain were processed through the artificial intelligence layer, evaluated within a digital twin environment, converted into feasible recovery decisions, and updated through feedback learning. The analytical sequence was defined as in Figure 1.

Figure 1 presents the proposed closed-loop supply-chain intelligence framework. Operational data are captured from the physical supply chain and processed through the data layer for predictive analytics. The AI prediction engine identifies disruption risks, while the digital twin evaluates alternative recovery scenarios. Risk evaluation supports informed decision-making, and feedback learning continuously updates the system, enabling adaptive resilience and improved future performance.

2.2 Supply-chain system representation

The supply chain was represented as a directed network consisting of suppliers, production facilities, warehouses, distribution centres, transport routes, retailers, and final customers. Each node represented a decision point, while each edge represented a material, information, or logistics flow. The network was expressed as Olan et al. (2024):

G = ( N , E ) (1)

Where G is the supply-chain network, N is the set of operational nodes, and E is the set of flow links connecting the nodes.

For each time period t, the system state was represented by an operational feature vector:

X t = { L t , D t , I t , Q t , C t , A t , S t , R t } (2)

Where L_t is supplier lead time, D_t is demand, I_t is inventory level, Q_t is order quantity, C_t is logistics cost, A_t is route availability, S_t is service level, and R_t is recorded disruption status. The response variables were disruption probability, vulnerability score, expected delay, stock-out risk, recovery cost, and resilience index.

2.3 Data sources, dataset composition and synthetic augmentation

The study used a hybrid dataset comprising a de-identified operational component and a simulation-generated augmentation component. The operational component contained 3,500 observations, representing 70.00% of the complete analytical dataset, and was obtained from CarryGo Logistics Company, a logistics and supply-chain service company located in Awka, Anambra State, Nigeria. The records covered the period from January 2023 to December 2025 and represented individual shipment and delivery transactions handled through the company’s supplier, warehouse and transport network. Each observation corresponded to a traceable shipment transaction recorded at a defined time and linked to its originating supplier, handling warehouse, transport route and delivery outcome. The operational records were extracted from the company’s shipment registers, dispatch records, warehouse inventory logs, supplier delivery records, route-monitoring reports, transport-cost records and customer-delivery documentation. The variables obtained directly from these records included supplier identifier, warehouse identifier, route identifier, supplier lead time, demand volume, available inventory, order quantity, transport cost, route availability, service level, production or handling capacity, supplier reliability and recorded disruption status. Weather severity, port-congestion conditions and fuel-scarcity indicators were matched to the operational observations using the corresponding transaction date, shipment origin, shipment destination and transport corridor. Where exact daily values were unavailable, the indicators were assigned from the nearest verified reporting period applicable to the shipment location and date. To protect commercial confidentiality, the company name was retained only with organisational approval, while supplier names, customers, warehouses, drivers, vehicle numbers and route-specific commercial identifiers were replaced with non-identifying alphanumeric codes before analysis. The simulation-generated component contained 1,500 observations, representing 30.00% of the complete dataset. It was introduced to increase the representation of infrequent but operationally consequential disruption conditions that were not sufficiently represented in the company’s historical records. These included severe supplier delays, simultaneous route and fuel disruptions, substantial demand surges, port-related delays, abrupt inventory shortages and compound disruption events. The final analytical dataset therefore contained:

N = N R + N S = 3 , 500 + 1 , 500 = 5 , 000 (3)

Observations, where N R represents the operational component and N S represents the simulation-generated component. A binary variable named “Data Origin” was retained throughout data preparation and modelling, with “Data Origin” = 0 denoting a real operational observation and “Data Origin” = 1 denoting a simulation-generated observation. The variable was used for source tracking, data-quality checks and source-stratified validation, but it was excluded from the predictor set to prevent the machine-learning models from identifying disruption status through the origin of an observation. This procedure ensured that simulated records were not presented as directly observed industrial events and allowed model performance to be evaluated independently on genuine company data.

2.3.1 Distribution fitting and generation of synthetic observations

The simulation parameters were estimated from the 3,500 operational observations rather than assigned through unrestricted assumptions. Distribution fitting was undertaken separately for each numerical variable using the operational training records. Candidate probability distributions were compared using the Akaike information criterion, Bayesian information criterion, Kolmogorov–Smirnov goodness-of-fit statistic and visual comparison of the empirical and fitted distributions. The distribution with the lowest information criterion and an acceptable goodness-of-fit result was retained for each variable. Supplier lead time and transport cost were generated from fitted lognormal distributions because both variables were strictly positive and showed right-skewed behaviour in the operational records. Demand volume and order quantity were generated from fitted gamma distributions, which accommodated positive values, operational variability and occasional high-volume transactions. Supplier reliability, route availability and service level were bounded within the interval [0,1] and were therefore generated from fitted beta distributions. Count-based operational occurrences, including the number of delayed deliveries, stock-out incidents and route-interruption events within each reporting window, were generated using negative-binomial distributions because their variance exceeded their mean and therefore did not satisfy the equidispersion assumption of a standard Poisson model. Inventory level and production or handling capacity were generated through an empirical moving-block bootstrap from the operational records. This method was preferred to unrestricted parametric sampling because it preserved the actual ranges, serial dependence and joint operational patterns found in the company data. Seven-day blocks of consecutive records were sampled with replacement, and small random perturbations drawn from the corresponding within-block residual distribution were added to prevent exact duplication. Generated inventory and capacity values were truncated at zero and capped at the observed 99th percentile to prevent impossible negative quantities or unrealistic extreme values. Categorical identifiers, including supplier, warehouse and route codes, were sampled according to their observed frequencies in the operational training subset. Conditional dependencies were retained during sampling. For example, route identifiers were sampled within the applicable warehouse and destination combination, supplier reliability was conditioned on supplier category, and transport cost was conditioned on route distance, route availability and shipment volume. This prevented the generation of combinations that could not occur within the company’s operating network.

Temporal dependence was preserved through a combination of the seven-day moving-block bootstrap and first-order lag conditioning. Simulated demand, lead time, inventory level, transport cost and service level were conditioned on their immediately preceding operational values and their 7-day and 28-day rolling averages. Consequently, the simulation did not treat each observation as an independent random record. Instead, it retained short-term persistence, weekly operating patterns and the tendency of logistics disruptions to continue across neighbouring periods.

2.3.2 Disruption-injection rules

Disruption scenarios were introduced through explicit conditional rules derived from the distributional properties of the operational training data. The thresholds were estimated exclusively from the training period and were not recalculated using validation or test observations. A supplier-delay event was assigned when realised lead time exceeded the supplier-specific 90th percentile of normal lead time or when supplier reliability fell below its operational 10th percentile. Where both conditions occurred within the same observation, the event was treated as a severe supplier disruption. Supplier-delay injection increased baseline lead time by a randomly sampled value between 20 and 50% and reduced expected service level by 5 to 15 percentage points, depending on the severity of the reliability reduction. A route-disruption event was generated when route availability fell below the route-specific 15th percentile and at least one supporting external-risk condition was present. The supporting conditions were weather severity above its 75th percentile, fuel-scarcity index above its 75th percentile, or an existing route interruption recorded within the preceding seven-day window. A moderate route disruption reduced route availability by 15–30%, while a severe disruption reduced it by 31–60%. Generated route-availability values were constrained to the interval [0,1]. A port-congestion event was introduced where a shipment depended directly or indirectly on imported goods, port clearance or long-haul transfer from a port-linked corridor. Moderate congestion increased shipment lead time by 20–35% and transport cost by 15–25%. Severe congestion increased lead time by 36–60% and transport cost by 26–40%. The corresponding service level was reduced by 8 to 20 percentage points, subject to a lower bound of zero. A fuel-scarcity event was generated when the fuel-scarcity index exceeded its 75th percentile. The event increased transport cost by 10.00–30.00%, reduced route availability by 5–20%, and increased delivery lead time by 10–25%. The magnitude of the change was conditioned on route distance because long-distance movements were more exposed to fuel availability and refuelling constraints than short urban deliveries. A demand-surge event was defined as demand exceeding the corresponding 28-day rolling mean by at least 30%. Moderate demand surges increased the baseline demand by 30–49%, while severe surges increased it by 50–80%. Order quantity and inventory depletion were adjusted consistently with the generated demand level so that demand, order and stock variables remained operationally connected. An inventory-shortage event was assigned when available inventory was lower than the sum of forecast demand over the replenishment lead time and the applicable safety-stock requirement (Theissler et al., 2022):

I t < D ̂ L , t + S S t (4)

Where I t is the available inventory at time t, D ̂ L , t is forecast demand over the applicable replenishment lead time and S S t is safety stock. Safety stock was computed at a 95% service-protection level as:

S S t = 1.645 σ D L t (5)

Where σ D is the standard deviation of demand and L t is the corresponding replenishment lead time. An inventory shortage was considered severe where available inventory was below 70% of the combined forecast-demand and safety-stock requirement. A compound disruption was created where two or more disruption conditions occurred within the same shipment record or within the same seven-day operating window. Examples included supplier delay combined with fuel scarcity, port congestion combined with route unavailability, or demand surge occurring together with low inventory. Compound disruption scenarios were retained because the operational consequences of interacting events are usually more serious than the effect of an isolated event. The final disruption status was assigned using a rule-based severity score. Each moderate disruption condition contributed a score of 1, while each severe condition contributed a score of 2. An observation was labelled as disrupted where at least one severe disruption occurred or where the cumulative disruption score was equal to or greater than 2:

Y t = { 1 , if max ( s kt ) = 2 or ∑ k = 1 K s kt ≥ 2 0 , otherwise (6)

Where s kt ∈ { 0 , , 1 , , 2 } denotes the absence, moderate presence or severe presence of disruption type k at time t. The generated label was subsequently checked against the direction of change in lead time, service level, route availability, inventory adequacy and transport cost to ensure that a disrupted label was not assigned to an operationally normal observation.

2.3.3 Training, validation and testing structure

To prevent information leakage, the 3,500 operational observations were first arranged chronologically according to shipment date and divided before synthetic augmentation. The earliest 2,000 operational observations were assigned to the model-development period. The next 750 operational observations were reserved as the validation subset, while the most recent 750 operational observations were retained as the final test subset. The 1,500 simulation-generated observations were then added exclusively to the 2,000-observation operational training subset.

The resulting analytical structure was:

N training = 2 , 000 operational + 1 , 500 simulated = 3 , 500 (7)

N validation = 750 operational

N testing = 750 operational .

This produced a final 70:15:15 training, validation and testing ratio, comprising 3,500 training observations, 750 validation observations and 750 test observations. The training subset was used for parameter estimation, feature engineering, model fitting and class-imbalance adjustment. The validation subset was used for hyperparameter selection, probability-threshold calibration, early stopping and model comparison during development. The final test subset was used once for independent performance reporting, paired bootstrap analysis, McNemar-type comparison and operational generalisation assessment. No simulated observation was included in either the validation or final test subset. Furthermore, the 750 test observations were not used to fit probability distributions, estimate synthetic-data parameters, construct disruption thresholds, select model hyperparameters, determine classification cut-offs or calculate feature-normalisation statistics. All preprocessing parameters, including the mean, standard deviation, percentile thresholds and scaling limits, were estimated from the training subset and applied unchanged to the validation and test subsets. This structure allowed the models to benefit from synthetic rare-event augmentation while ensuring that the reported final performance was obtained entirely from unseen operational records collected from CarryGo Logistics Company (see Table 1).

Table 1. Composition and analytical allocation of the hybrid dataset.

2.4 Experimental setup and reproducibility protocol

The final hybrid dataset was divided into 3,500 training observations, 750 validation observations and 750 test observations, corresponding to a 70:15:15 ratio. The training subset combined 2,000 chronologically earlier operational records with 1,500 simulation-generated rare-disruption records. The validation and test subsets each contained 750 subsequent operational records from CarryGo Logistics Company and did not contain any simulated observation. The validation subset supported model selection and threshold calibration, while the final operational test subset was used exclusively for independent performance reporting.

2.4.1 Digital twin development

The digital twin was developed as a dynamic virtual representation of the physical supply-chain network. It captured node capacity, inventory position, supplier reliability, transport constraints, demand behaviour, route availability, disruption events, and recovery actions. The twin state at time t + 1 was updated from the current state, selected action, and disruption condition as Ivanov (2020):

S t + 1 = f ( S t , U t , Δ t ) (8)

where S_t is the supply-chain state at time t, U_t is the selected recovery or control action, and Δ t is the disruption condition. The twin was used to mirror current performance, simulate disruption scenarios, and evaluate recovery strategies before implementation. The digital-twin transition function was implemented as a deterministic state-update mechanism with disruption and recovery inputs:

S t + 1 = f ( S t , U t , δ t ) (9)

where:

S t = [ I t , B t , L t , A t , S L t , C t ] ⊤ (10)

represents inventory, backorders, lead time, route availability, service level and cumulative cost. The principal state updates were:

I t + 1 = max ( 0 , I t + Q t in − Q t out − D t ) (11)

B t + 1 = max ( 0 , B t + D t − I t − Q t in )

L t + 1 = L t 0 ( 1 + δ t sup + δ t port + δ t weather + δ t fuel ) − U t L

A t + 1 = min [ 1 , max ( 0 , A t − δ t route − δ t weather − δ t fuel + U t A ) ]

S L t + 1 = 1 − B t + 1 D t + ε

where U t L and U t A represent the lead-time and route-availability improvements produced by the selected recovery action, and εis a small constant preventing division by zero.

Propagation logic: Disruption propagation from upstream node i to downstream node j was represented as:

δ j , t + 1 = min [ 1 , max ( 0 , δ j , t + ρ ij δ i , t ( 1 − R C j , t ) − λ j U j , t ) ] (12)

Where ρ ij is the disruption-transfer coefficient, R C j , t is the resilience capacity of the receiving node, and λ j U j , t represents disruption reduction achieved through recovery action. Propagation occurred only across active network links and continued until disruption severity fell below the intervention threshold. Simulated outputs were compared with observed outcomes using prediction error:

e t = Y t − Y ̂ t (13)

Where Y_t is the observed supply-chain performance outcome and Ŷ_t is the digital twin-simulated outcome.

2.5 Machine-learning prediction and model validation

Machine-learning models were trained to predict disruption vulnerability and supply-chain performance degradation. Random Forest and XGBoost were applied to structured operational data because they can capture nonlinear relationships and interaction effects among supplier reliability, inventory level, transport performance and demand variability. LSTM was applied to sequential variables such as demand trend, lead-time behaviour and delay pattern, while Bayesian Neural Network modelling was used where probabilistic prediction and uncertainty estimation were required. To reduce data leakage, a time-aware validation strategy was applied. Earlier operational records were used for model training, while later observations were reserved for validation and testing. Model selection was based on combined classification and regression performance, with priority given to disruption-detection sensitivity, AUC stability and prediction-error reduction. This validation approach ensured that model performance reflected future-facing disruption prediction rather than random mixing of past and future supply-chain observations. The disruption prediction model was specified as Arrieta et al. (2020):

p ̂ t = P ( Y t = 1 ∣ X t ) (14)

Where p_t is the predicted probability of disruption at time t, Y_t = 1 denotes a disrupted state, and X_t is the operational feature vector. Regression outputs such as expected delay, recovery time, and recovery cost were modelled as:

Y ̂ t = f ( X t ) + ε t (15)

where Y ̂ t is the predicted response and ε _ t is the random error term. Classification models were evaluated using accuracy, precision, recall, F1-score, and AUC. Regression models were evaluated using RMSE, MAE, and R². RMSE was computed as:

RMSE = 1 n ∑ i = 1 n ( Y i − Y ̂ i ) 2 (16)

2.5.1 Paired bootstrap and McNemar-type statistical validation

To further assess the stability of the comparative model evaluation, paired bootstrap resampling and McNemar-type comparative analyses were performed after the primary performance assessment. These procedures were used to determine whether the observed differences among Random Forest, XGBoost, Long Short-Term Memory (LSTM), and Bayesian Neural Network (BNN) models remained consistent beyond point estimates of accuracy, precision, recall, F1-score, AUC, RMSE, and MAE. Paired bootstrap validation was implemented by repeatedly resampling the independent test-set evaluation context with replacement over 10,000 iterations. The distribution of pairwise accuracy differences was used to estimate mean performance differences and corresponding 95% confidence intervals. Model comparisons were considered practically stable when the confidence interval excluded zero. McNemar-type comparative analysis was subsequently performed to examine pairwise classification disagreement between competing models using discordant prediction outcomes. This analysis evaluated whether the observed differences in classification behaviour were statistically meaningful rather than resulting from sampling variation. The combined application of paired bootstrap resampling and McNemar-type comparison provided an additional layer of statistical robustness for comparative model assessment and reduced the likelihood of overinterpreting small differences in predictive performance. These complementary analyses therefore strengthened the reliability of the disruption-prediction evaluation and improved confidence in the comparative conclusions reported in this study.

2.5.2 Validation on held-out operational data

The predictive models were independently validated using operational records obtained from CarryGo Logistics Company, Awka, Anambra State, Nigeria. The validation subset contained 750 chronologically later operational observations, while the final test subset contained another 750 unseen operational observations. No simulation-generated observation was included in either subset. The simulated records were used only to augment the training dataset and improve the representation of rare disruption events. The validation subset was used for model selection, hyperparameter tuning and probability-threshold calibration. The final operational test subset was retained for independent reporting of accuracy, precision, recall, F1-score, AUC, RMSE and MAE. All preprocessing parameters, disruption thresholds and synthetic-data distributions were estimated from the training data and applied unchanged to the operational validation and test subsets. This procedure ensured that the reported model performance reflected prediction on genuine company operations rather than performance on simulation-generated cases.

2.6 Vulnerability and resilience assessment

Supply-chain vulnerability was estimated from predicted disruption probability, expected delay, stock-out risk, service-level loss, and recovery cost. The composite vulnerability score was defined as Queiroz et al. (2022):

V t = 0.20 P ˜ t + 0.20 D ˜ t + 0.20 SO ˜ t + 0.20 SL ˜ t + 0.20 RC ˜ t (17)

Where V_t is the vulnerability score, P_t is disruption probability, D_t is expected delay, SO_t is stock-out risk, SL_t is service-level loss, RC_t is recovery cost, and w₁ to w₅ are normalized weights. Thus:

w 1 = w 2 = w 3 = w 4 = w 5 = 0.20 , ∑ k = 1 5 w k = 1 (18)

Equal weights were used to avoid assigning unsupported priority to any vulnerability component.

For risk-oriented variables, min–max normalization was applied as:

x ˜ t = x t − x min x max − x min + ε (19)

Because higher service level indicates lower vulnerability, service-level loss was normalized as:

SL ˜ t = 1 − S L t − S L min S L max − S L min + ε (20)

All minimum and maximum values were estimated from the training subset and applied unchanged to validation and test data.

The resilience index was computed to reflect the ability of the supply chain to maintain service performance and recover after disruption:

R I t = S t S 0 × 1 1 + T r (21)

Where RI_t is the resilience index, S_t is service level after disruption, S₀ is baseline service level, and Tᵣ is recovery time. A higher value indicates stronger resilience.

2.7 Uncertainty quantification and decision-confidence assessment

Uncertainty was quantified using Bayesian Neural Networks and Monte Carlo simulation. The Bayesian Neural Network generated probabilistic disruption predictions by estimating a distribution of possible outputs rather than a single deterministic value. Monte Carlo simulation was then used to repeatedly evaluate disruption outcomes under varying assumptions of lead time, demand fluctuation, supplier reliability, route availability, transport delay and recovery cost. A total of 10,000 Monte Carlo iterations was used to estimate expected-case, best-case and worst-case outcomes for delay duration, recovery cost and service-level loss. For each scenario, predictive means, uncertainty bands and confidence intervals were computed. Uncertainty estimates were interpreted as decision-confidence indicators. Predictions with high disruption probability and narrow uncertainty intervals were treated as intervention-ready cases, while high-risk predictions with wide uncertainty intervals were flagged for additional monitoring before costly recovery actions were implemented. This procedure ensured that recovery decisions were not based only on predicted risk magnitude, but also on the reliability of the prediction.

The Bayesian prediction was represented as:

P ( Y ∣ X , D ) = ∫ P ( Y ∣ X , θ ) P ( θ ∣ D ) dθ (22)

Where P ( Y ∣ X , D ) is the posterior predictive distribution, θ represents model parameters, and D is the training dataset.

2.8 Explainable artificial intelligence using SHAP

SHAP analysis was used to explain the contribution of each input variable to disruption vulnerability and recovery outcomes. Global SHAP analysis was applied to rank the dominant disruption drivers across the full test dataset, while local SHAP analysis was used to explain individual high-risk disruption cases. TreeSHAP was applied to the Random Forest and XGBoost models because these models produced the strongest predictive performance and are compatible with efficient tree-based explanation. The SHAP outputs were interpreted at two levels. First, mean absolute SHAP values were used to rank the overall importance of variables such as supplier reliability, lead time, inventory level, order quantity, transport cost, demand volume, route availability, weather severity, port congestion and fuel scarcity index. Second, local SHAP values were used to identify whether each variable increased or reduced the predicted risk for a specific disruption case. This made the prediction layer transparent and managerially interpretable. For each prediction, the SHAP value decomposed model output into feature-level contributions:

f ( x ) = ϕ 0 + ∑ j = 1 m ϕ j (23)

Where f(x) is the model prediction, ϕ₀ is the baseline prediction, ϕj is the contribution of feature j, and m is the number of input features. This enabled the framework to explain whether disruption risk was mainly caused by supplier delay, low inventory, demand surge, route unreliability, high logistics cost, port delay, or capacity limitation.

2.9 Adaptive recovery strategy evaluation and optimization

The adaptive decision layer selected recovery actions after prediction, uncertainty assessment, explainability analysis and digital twin simulation. Candidate strategies included supplier substitution, order reallocation, inventory redistribution, route diversion, emergency procurement, production rescheduling, safety-stock adjustment and transport-mode substitution. Each strategy was evaluated against four decision criteria: recovery cost, delay reduction, service-level preservation and disruption-risk reduction. Recovery strategies were ranked using a multi-objective trade-off logic. The selected strategy was not necessarily the cheapest option, but the alternative that achieved the best balance between resilience performance and operational feasibility. Equal initial weights were assigned to the four optimization criteria to avoid subjective bias. Sensitivity checks were then used to confirm whether the preferred strategy remained stable when higher priority was assigned to service continuity or risk reduction. The optimization objective was to minimize disruption risk, delay, recovery cost, and service-level loss. The recovery objective was defined as:

J ( U ) = α 1 C ˜ ( U ) + α 2 [ 1 − DR ˜ ( U ) ] + α 3 [ 1 − SL ˜ ( U ) ] + α 4 [ 1 − RR ˜ ( U ) ] (24)

Where C is recovery cost, DR is delay reduction, SL is service level and RR is disruption-risk reduction.

The baseline weights were:

α 1 = α 2 = α 3 = α 4 = 0.25 , ∑ k = 1 4 α k = 1 (25)

Each criterion was normalized as:

x ˜ i = x i − x min x max − x min + ε (26)

For beneficial criteria (delay reduction, service level and risk reduction) the complement 1 − x ˜ i was used so that every term represented a loss to be minimized. Therefore, the recovery strategy with the lowest J ( U ) was ranked as the preferred alternative.

2.10 Feedback learning and model updating

After each implemented recovery action, actual outcomes were compared with predicted and simulated outcomes. The resulting error was used to update the prediction model and recalibrate the digital twin. The feedback error was defined as:

F E t = Y t − Y ̂ t DT (27)

Where FE_t is the feedback error, Y_t is the observed outcome, and Ŷ_tᴰᵀ is the digital twin-simulated outcome. Model updating was triggered when feedback error or prediction drift exceeded a predefined tolerance. This ensured that the framework remained adaptive under changing supply-chain conditions and did not depend on static historical assumptions. The feedback-learning process was implemented by comparing observed post-decision outcomes with predicted and simulated outcomes. When the feedback error exceeded the predefined tolerance threshold, the model was updated using the most recent operational records and the digital twin state variables were recalibrated. This process enables periodic model recalibration using newly available operational data, allowing the framework to adapt to changing supply-chain conditions while maintaining prediction reliability. The feedback loop therefore converted operational errors into model-improvement signals, making the framework adaptive rather than static.

3 Results and discussion

3.1 Digital twin representation of supply-chain operations

Figure 2 illustrates the transformation of physical supply-chain operations into a connected digital twin environment, thereby addressing Objective 1. The representation of suppliers, manufacturing facilities, warehouses, distribution centres, retailers, and customers confirms that the developed model captures multi-echelon supply-chain interactions. The network structure further reveals possible disruption transmission pathways across operational nodes, emphasizing the importance of system-wide visibility instead of isolated monitoring. This supports the need for integrated resilience assessment across the entire supply network.

Figure 3 demonstrates the data-driven foundation required for intelligent supply-chain decision support. The integration of ERP records, procurement information, inventory data, supplier performance reports, demand patterns, transport logs, weather information, port activities, and disruption histories confirms comprehensive monitoring of internal and external risk indicators. This architecture strengthens methodological reliability by ensuring that prediction, simulation, and optimization processes are supported by diverse operational evidence rather than theoretical assumptions, thereby reinforcing the digital twin development objective.

Figure 4 captures the dynamic synchronization between operational variables and the virtual supply-chain environment. The interaction among inventory status, lead time, demand behaviour, transportation performance, service level, and disruption condition confirms that the digital twin continuously reflects changing system states. This capability is critical because effective resilience management depends on real-time adaptation rather than static analysis. The synchronization mechanism justifies the application of dynamic modelling for disruption monitoring, early-warning prediction, and adaptive decision support.

Table 2 presents the analytical foundation linking supply-chain data characteristics with model development requirements. The classification of operational variables, performance indicators, supplier attributes, cost parameters, external risk factors, and target outputs confirms strong alignment between data structure and research objectives. The inclusion of disruption status enables supervised learning implementation, while variables such as supplier reliability, service level, and lead time ensure adequate representation of resilience behaviour. This strengthens the methodological traceability of the proposed framework.

Table 2. Supply-chain variables and model inputs.

3.2 Prediction of disruption vulnerability using machine learning models

Figure 5 describes the computational pathway adopted for transforming supply-chain data into predictive intelligence. The sequence from preprocessing, feature engineering, Random Forest, XGBoost, LSTM modelling, prediction generation, and disruption classification confirms a structured machine-learning implementation process. The workflow improves methodological transparency by showing how raw operational records are systematically converted into actionable early-warning outputs. This validates the integration of artificial intelligence as the predictive engine within the digital twin framework.

Table 3 shows that Random Forest achieved the highest disruption-detection sensitivity, with an accuracy of 0.79, recall of 0.99 and AUC of 0.85, making it the most effective model for identifying potential disruption events. However, its precision of 0.61 indicates that some predicted disruption alerts may not correspond to actual disruption cases. In contrast, XGBoost achieved a slightly higher precision of 0.62, the same AUC of 0.85, and the lowest MAE of 0.27, providing a more balanced prediction performance. A precision of 0.61 means that approximately 39% of Random Forest disruption alerts did not correspond to observed disruption cases. Such false alerts may lead to unnecessary supplier escalation, route review, emergency procurement or inventory redistribution. Random Forest should therefore be used as a high-sensitivity screening model, with uncertainty assessment and managerial verification applied before costly recovery actions are implemented. These findings suggest that Random Forest is better suited for high-sensitivity early-warning applications where missing a disruption is more costly than investigating additional alerts, while XGBoost offers a suitable alternative when reducing unnecessary intervention is a greater operational priority.

Table 3. Performance comparison of machine learning models.

Figure 6 reveals the effectiveness of the prediction engine in categorizing supply-chain operating conditions into actionable vulnerability levels. The classification identified 353 normal states, 61 watch states, 333 warning states, and 3 critical conditions. The substantial number of warning observations highlights the existence of hidden operational vulnerabilities before severe disruption occurs. This outcome confirms the framework’s early-warning capability and demonstrates its usefulness for proactive intervention instead of delayed disruption response.

3.2.1 Statistical robustness of model comparisons

Figure 7 strengthens the model-comparison evidence by showing the resampled distribution of paired accuracy differences across the evaluated models. The bootstrap results indicate that Random Forest was statistically stronger than LSTM, with an observed accuracy difference of 0.050667 and 95% CI of 0.021333–0.080033, and also stronger than BNN, with difference of 0.030667 and 95% CI of 0.002667–0.06. XGBoost also outperformed LSTM significantly. However, Random Forest and XGBoost were not statistically different. This means the ensemble models are dependable, but RF’s superiority over XGBoost should not be overclaimed.

Figure 8 provides a -level statistical summary by plotting observed accuracy differences with 95% bootstrap confidence intervals. The figure confirms that significant model separations occur only where the confidence interval does not cross zero. Random Forest versus LSTM, Random Forest versus BNN, and XGBoost versus LSTM show statistically stable differences. In contrast, Random Forest versus XGBoost, XGBoost versus BNN, and BNN versus LSTM cross zero, showing non-significant differences.

Table 4 strengthens the comparative evaluation of the machine-learning models by assessing whether the observed accuracy differences remain stable under repeated paired bootstrap resampling. Random Forest demonstrated statistically stable improvements over both LSTM and Bayesian Neural Network because their corresponding 95% confidence intervals excluded zero, indicating consistent performance superiority across repeated resampling. XGBoost also showed a significant advantage over LSTM, whereas the comparison between Random Forest and XGBoost remained statistically indistinguishable because the confidence interval crossed zero. These findings indicate that although Random Forest achieved the highest predictive performance, its advantage is principally established over weaker models rather than over XGBoost. Consequently, the results provide stronger statistical support for selecting Random Forest while avoiding exaggerated claims of universal superiority.

Table 4. Paired bootstrap accuracy comparison.

Figure 9 directly examines paired classification disagreement between model pairs using discordant correctness counts. Random Forest versus LSTM produced 85 cases correctly classified by Random Forest but wrongly classified by LSTM, against 47 in the opposite direction, confirming RF’s stronger classification consistency. XGBoost versus LSTM showed a similar advantage, with 74 against 44 discordant cases. However, Random Forest versus XGBoost had closer discordance counts, 45 against 37, showing no convincing superiority. This figure is valuable because it moves the manuscript beyond aggregate accuracy and tests whether model differences are practically consistent on the same test observations.

Figure 10 confirms which model-pair differences remain statistically meaningful under McNemar-type exact testing. Random Forest significantly outperformed LSTM, with exact p = 0.001195, while XGBoost also significantly outperformed LSTM, with exact p = 0.007330. However, Random Forest versus XGBoost was not significant, with exact p = 0.439705. Random Forest versus BNN was borderline but not significant by exact testing, p = 0.050487. This result supports a careful interpretation: RF is the strongest disruption detector, but its advantage is statistically decisive mainly against LSTM, not against XGBoost.

Table 5 evaluates whether differences in classification performance arise from systematic improvements or merely from random disagreement between competing models. The analysis shows that Random Forest significantly outperformed LSTM, while XGBoost also achieved significantly better classification performance than LSTM, confirming the robustness of the ensemble-learning approaches. In contrast, the comparison between Random Forest and XGBoost was not statistically significant, indicating that both models produced broadly comparable classification decisions on the evaluated disruption cases. The borderline result observed between Random Forest and Bayesian Neural Network further suggests only modest practical differences. Overall, the McNemar-type comparison strengthens the research by demonstrating that the reported model rankings are supported by formal comparative statistical assessment rather than by descriptive performance metrics alone.

Table 5. McNemar-type classification comparison.

3.3 Digital twin simulation of disruption propagation

Figure 11 provides quantitative evidence that the proposed digital twin accurately represents disruption propagation across interconnected supply-chain stages. The simulation demonstrates a clear escalation in disruption severity, increasing from supplier failure (22%) to production delay (36%), inventory shortage (51%), distribution delay (68%), and ultimately customer service loss (83%). This systematic progression confirms that localized upstream disruptions propagate throughout the network and intensify downstream operational consequences when timely mitigation measures are absent. The observed amplification validates the digital twin’s capability to model complex interdependencies and accurately reproduce cascading disruption behaviour within multi-echelon supply chains. Beyond visualization, the simulation provides a robust analytical basis for identifying critical vulnerability points, evaluating disruption trajectories, and assessing the likely consequences of delayed interventions before they occur in practice. These capabilities strengthen the framework’s value as a predictive decision-support tool for resilience planning and adaptive recovery. Nevertheless, although the simulation results demonstrate strong internal validity, future research should further evaluate the framework using benchmark digital twin platforms, historical industrial disruption events, and live operational supply-chain data to establish broader empirical generalisability and industrial applicability.

3.4 Uncertainty quantification and decision confidence assessment

Figure 12 evaluates uncertainty behaviour by simulating multiple disruption conditions affecting delay duration, recovery cost, and service-level performance. The distribution of best-case, expected-case, and worst-case scenarios demonstrates that supply-chain decisions cannot rely only on deterministic predictions. The simulation enhances decision reliability by revealing possible operational variations before implementation. This supports uncertainty-aware resilience management where recovery strategies are selected based on expected performance and associated risk exposure.

Figure 13 illustrates the variability associated with disruption predictions through confidence-bound analysis. Changes in prediction intervals indicate that some operational conditions produce highly reliable forecasts, whereas others require cautious interpretation due to greater uncertainty. This distinction strengthens decision governance by preventing excessive dependence on single prediction values. The uncertainty assessment confirms that the framework can evaluate both disruption likelihood and prediction confidence, improving reliability under unstable supply-chain conditions.

3.5 Identification of dominant disruption drivers using explainable artificial intelligence

Figure 14 identifies the dominant operational factors influencing disruption vulnerability through explainable artificial intelligence. Supplier reliability and lead-time days produced the strongest influence, with approximate mean SHAP values of 1.18 and 1.14, respectively, followed by inventory level at 0.18. These results confirm that disruption risk is primarily influenced by supplier performance and delivery reliability. The explainability output converts machine-learning predictions into actionable managerial insights by identifying priority areas for resilience improvement.

Figure 15 presents a local TreeSHAP explanation generated from a representative unseen operational test observation. The figure quantifies the contribution of each predictor to the model’s disruption prediction using SHAP values computed directly from the trained Random Forest model. The explanation reveals how the combined effects of the operational variables influence the final prediction, providing interpretable evidence of the decision-making process and supporting confidence in the model’s application to supply-chain disruption assessment.

Table 6 translates SHAP-based explanations into measurable disruption priorities. Supplier reliability emerged as the most influential factor with mean SHAP value of 1.18 and 41.49% contribution, followed by lead time with 1.14 and 39.97%. Inventory level accounted for 6.34%, whereas other factors contributed below 2.00% individually. The ranking confirms that supplier stability and delivery performance represent the most critical intervention areas for improving supply-chain resilience.

Table 6. Ranking of dominant disruption drivers.

3.6 Adaptive recovery strategy evaluation and optimization

Figure 16 compares the five recovery strategies using the normalized quantitative criteria reported in Table 7. Supplier switching produced the strongest cost and delay performance, while inventory redistribution provided the highest service level and the lowest equal-weight objective value. Production rescheduling achieved the greatest risk reduction but performed less favourably on recovery cost and delay. The figure therefore illustrates the trade-offs that produced the final multi-objective ranking.

Table 7. Multi-objective optimization results.

Table 7 shows that inventory redistribution provided the best overall recovery balance, recording the lowest objective value of 0.26, service level of 0.81, delay reduction of 1.60%, and risk reduction of 49.23%. Supplier switching ranked second with an objective value of 0.39, supported by the lowest recovery cost of 25,002.70 and highest delay reduction of 4.38%. Route diversion ranked third but recorded the lowest risk reduction of 45.33%. Emergency procurement ranked fourth with an objective value of 0.6816; its 50.48% risk reduction was weakened by −3.62% delay reduction. Production rescheduling ranked fifth despite achieving 52.39% risk reduction because it combined the highest recovery cost of 26,530.54, −2.34% delay reduction, and no service-level advantage under the equal-weight multi-objective decision framework applied.

Figure 17 demonstrates how predictive outputs are converted into practical recovery decisions through simulation, risk assessment, optimization, recommendation, implementation, and feedback stages. The structured pathway confirms that resilience actions are evaluated before execution rather than selected reactively. This decision logic enhances practical applicability because managers can compare intervention options under simulated conditions before committing resources, thereby supporting adaptive and evidence-based supply-chain recovery.

Table 8 confirms that recovery performance changes with managerial priorities. Under equal weights, inventory redistribution ranked first with J = 0.26, followed by supplier switching at 0.39, route diversion at 0.67, emergency procurement at 0.68, and production rescheduling at 0.71. Under cost priority, supplier switching became first at 0.24, while inventory redistribution followed at 0.25. The same pattern appeared under delay priority, where supplier switching recorded 0.24 and inventory redistribution 0.29. Under service priority, inventory redistribution remained best at 0.15, ahead of supplier switching at 0.64. Under risk priority, inventory redistribution ranked first at 0.33, while production rescheduling improved to second at 0.43. Overall, inventory redistribution showed the strongest balance, while supplier switching performed best when cost or delay dominated.

Table 8. Objective-function values and weight-sensitivity.

3.7 Feedback-enabled capability and evolution of supply-chain resilience

Figure 18 highlights the feedback-enabled capability achieved through continuous comparison between predicted and observed outcomes. Prediction errors are transformed into learning information for model recalibration and digital twin improvement. This mechanism prevents system degradation caused by changing supply-chain conditions and ensures continuous adaptation. The feedback architecture confirms the major contribution of the study by enabling supply chains to evolve from static monitoring systems into continuously improving intelligent systems.

3.8 Implications for cognitive supply-chain resilience

Figure 19 synthesizes the overall scientific contribution by contrasting conventional resilience approaches with the proposed intelligent framework. The movement from monitor–respond–recover toward sense–predict–simulate–adapt–learn demonstrates a fundamental shift from reactive management to proactive resilience intelligence. The framework integrates digital twin modelling, machine learning, explainability, uncertainty evaluation, optimization, and feedback mechanisms into a unified adaptive system, confirming its potential for next-generation cognitive supply-chain management.

4 Conclusion

4.1 Summary of key findings

This study developed and evaluated a feedback-enabled digital twin and explainable artificial intelligence framework for predictive resilience and adaptive decision-making in disrupted supply chains. The first objective was achieved through the construction of a multi-echelon digital twin that represented suppliers, manufacturing plants, warehouses, distribution centres, retailers, and customers as interconnected operational nodes. This structure confirmed that supply-chain resilience cannot be adequately assessed at isolated points because disruption behaviour is transmitted through network relationships. The second objective was achieved by embedding machine-learning prediction into the digital twin environment. The predictive layer classified operational states into normal, watch, warning, and critical conditions, showing that a large proportion of supply-chain vulnerability occurs before full disruption becomes visible. This finding is technically important because it confirms the need for early-warning intelligence rather than post-event response. The third objective was achieved through uncertainty-aware simulation. Monte Carlo and probabilistic outputs demonstrated that recovery decisions vary under changing assumptions of delay, cost, service loss, supplier reliability, and demand fluctuation. This confirms that supply-chain decisions should not be based on deterministic estimates alone. The fourth objective was achieved through SHAP-based explainability. The analysis identified supplier reliability and lead-time behaviour as the dominant contributors to disruption vulnerability, while inventory position acted as a secondary but operationally significant driver. This confirms that disruption risk is not random but largely traceable to supplier-side dependability and delivery-time instability. The fifth objective was achieved through adaptive recovery evaluation and optimization. The recovery strategy assessment showed that inventory redistribution and supplier switching were the most competitive options, but for different reasons. Inventory redistribution strengthened service continuity and resilience protection, while supplier switching improved responsiveness and cost efficiency. This confirms that resilient decision-making requires trade-off optimization rather than selecting recovery strategies from a single performance measure. Random Forest is suitable where the cost of missing a disruption exceeds the cost of investigating an alert. However, its precision of 0.61 does not support the automatic execution of costly recovery actions without secondary verification. XGBoost provides a more balanced alternative where false-positive intervention costs are material.

4.2 Contributions

The study makes four major contributions to digital supply-chain resilience research. First, it presents an integrated closed-loop architecture that connects physical supply-chain operations, data integration, AI prediction, digital twin simulation, risk evaluation, adaptive decision-making, and feedback learning. This extends existing resilience models from descriptive monitoring to predictive and self-improving intelligence. Second, the study contributes a practically interpretable digital twin structure for disruption-prone supply chains. By modelling the supply chain as a dynamic network of operational nodes and flows, the framework supports real-time state synchronization and disruption-propagation analysis. Third, the study contributes an explainable prediction layer that links machine-learning outputs to managerial action. The SHAP analysis converts model predictions into ranked disruption drivers, thereby improving transparency and reducing the risk of black-box decision-making. Fourth, the study contributes an adaptive recovery logic that evaluates alternative actions through simulation before implementation. This is important because recovery decisions in unstable environments must balance cost, delay reduction, service preservation, risk reduction, and feasibility rather than depend on intuition or delayed manual response.

4.3 Comparison with previous studies

Previous supply-chain resilience studies have contributed useful knowledge on disruption monitoring, risk identification, recovery planning, digital twin modelling, and artificial intelligence applications (Bhandal et al., 2022; Nwamekwe and Igbokwe, 2024). However, many of these studies still approach resilience as a largely sequential activity, where firms detect disruption after operational performance has already started declining. In such models, the emphasis is often on response and recovery rather than early prediction, explanation, simulation, and continuous learning. This creates a practical gap, especially in Nigerian and similar disruption-prone environments where fuel scarcity, port congestion, supplier delay, road insecurity, and logistics instability can quickly spread across the entire supply chain. The present study improves on these earlier approaches by integrating prediction, digital twin simulation, explainable artificial intelligence, uncertainty analysis, recovery optimization, and feedback learning within one closed-loop framework (Olan et al., 2024; Kritzinger et al., 2018). Unlike conventional monitor–respond–recover models, the proposed framework follows a sense–predict–simulate–adapt–learn logic (Wang et al., 2022; Busse et al., 2021). This makes resilience more proactive, because the system does not wait for disruption to become severe before recommending action. It predicts vulnerability, explains the main risk drivers, evaluates uncertainty, tests recovery strategies, and updates itself using feedback from observed outcomes. This study also extends prior machine-learning supply-chain studies that focus mainly on predictive accuracy (Nwamekwe and Igbokwe, 2024). In this work, prediction is treated as only one part of decision intelligence. The model output is strengthened with SHAP-based explanation, Monte Carlo uncertainty assessment, digital twin scenario testing, and multi-objective recovery evaluation. Similarly, the digital twin is not used only as a virtual representation of the supply chain. It functions as a decision-governance tool for testing disruption propagation, comparing recovery options, and improving future decisions. Therefore, the major contribution of this study is that it moves supply-chain resilience from isolated digital tools to a more practical, explainable, adaptive, and feedback-enabled decision-support system suitable for unstable industrial environments.

4.4 Managerial and operational implications for predictive supply-chain resilience

The findings have direct implications for supply-chain managers, manufacturers, logistics providers, distributors, and policy-sensitive industries operating in uncertain environments. For Nigerian supply chains, where logistics delays, fuel scarcity, port congestion, supplier inconsistency, infrastructure constraints, insecurity, and exchange-rate volatility frequently affect operations, the framework offers a practical pathway for proactive decision support. The results show that firms should give priority to supplier reliability, lead-time stability, inventory visibility, and route performance because these factors strongly influence disruption exposure. Managers can use the framework to detect weak signals before disruption becomes severe, simulate recovery options, and select the most suitable intervention based on service continuity, cost, speed, and risk reduction. The study also supports the adoption of digital dashboards and decision-support systems that combine predictive analytics with human oversight. Instead of waiting for stock-outs, delayed deliveries, or customer service failure, organizations can monitor risk levels, evaluate uncertainty, and act before downstream losses accumulate. This is particularly valuable for industries where supply failure can affect production continuity, customer satisfaction, and financial performance.

4.4.1 Implementation requirements and deployment challenges

For real-world deployment, the proposed framework requires structured data pipelines linking ERP systems, procurement records, warehouse inventory logs, supplier performance reports, transport records, demand histories, port-clearance information, and disruption incident logs. It also requires a digital twin repository, machine-learning infrastructure, dashboard visualization, decision-rule configuration, and periodic model recalibration. Practical challenges may include incomplete data, inconsistent timestamp records, poor integration between departments, limited access to supplier data, computational cost, staff training needs, cybersecurity concerns, and managerial resistance to algorithm-supported decisions. In Nigerian supply-chain environments, additional challenges may arise from unstable logistics infrastructure, fuel scarcity, port delays, and unreliable real-time data availability. Therefore, phased implementation is recommended, beginning with high-risk supply-chain nodes before expanding into full network-wide deployment.

4.5 Limitations and future research

Although the framework was evaluated using a hybrid dataset, the final validation and testing were conducted exclusively on unseen operational records from CarryGo Logistics Company. This provides evidence that the predictive relationships transfer to real logistics operations. Nevertheless, the operational data originated from one company in Awka, Anambra State, Nigeria. Further validation across additional logistics firms, manufacturing supply chains and geographical locations is required before wider industrial generalisation. First, the study considered selected operational variables such as lead time, demand, inventory level, supplier reliability, transport cost, route availability, service level, and disruption status. These variables are relevant, but real supply chains in Nigeria and similar environments are affected by wider conditions. Future research should include richer data streams such as vehicle tracking, IoT sensor readings, port dwell-time records, supplier financial risk, exchange-rate volatility, geopolitical events, energy-price changes, and security-related disruptions. Second, the recovery optimization layer evaluated a limited set of recovery strategies. Further studies should include carbon emissions, contractual restrictions, sustainability targets, warehouse-capacity limits, labour constraints, customer-priority rules, and multi-period recovery planning. This will make the optimization model more realistic for industrial decision-making. Third, the feedback-learning mechanism requires stronger validation under live operational conditions. Future work should incorporate live model recalibration, human-in-the-loop governance, cybersecurity assessment, regulatory monitoring, and dashboard-based decision tracking. In practical Nigerian settings, deployment may also face data-quality gaps, weak system integration, unstable logistics infrastructure, staff training needs, and management resistance to algorithm-supported decisions. Addressing these issues will move the framework from analytical validation to full industrial adoption, while strengthening its value as a predictive, explainable, adaptive, and feedback-enabled supply-chain resilience system.

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.

Footnotes

Edited by: Fouad Jawab, Sidi Mohamed Ben Abdellah University, Morocco

Reviewed by: Riya Widayanti, Universitas Esa Unggul, Indonesia

Sanchari Ghosh, Indian Institute of Technology Dhanbad, India