Digital Twin-Driven Optimization of Mechanical Manufacturing Systems Using Artificial Intelligence

Digital Twin-Driven Optimization of Mechanical Manufacturing Systems  Using Artificial Intelligence

1.    Introduction
1.1 Background and Research Motivation
Manufacturing systems are undergoing a major technological transformation driven by Industry 4.0, cyber–physical systems, industrial Internet of Things (IIoT), cloud and edge computing, machine learning, and Digital Twin technologies[1,2,3]. The conventional manufacturing paradigm is increasingly being replaced by connected production environments in which physical equipment continuously generates data that can be analyzed to support monitoring, prediction, optimization, and control[4].
Among these technologies, the Digital Twin has become an important framework for linking physical manufacturing assets with dynamic virtual representations[5]. Rather than functioning as a static computer model, a manufacturing Digital Twin can continuously represent the operational state of machines and production processes and can be used to evaluate alternative operating conditions without directly interrupting physical production[6]. Recent research identifies manufacturing applications including process optimization, quality control, predictive maintenance, production scheduling, and system reconfiguration[7,8,9].
The integration of Artificial Intelligence further extends the functionality of Digital Twins[10]. Machine-learning algorithms can analyze sensor signals, historical production data, machine states, process parameters, and quality information to detect abnormal conditions and predict future system behavior[11]. Consequently, the Digital Twin can move beyond visualization and simulation toward predictive and prescriptive manufacturing[12]. Recent literature describes this transition as the development of AI-driven Digital Twins capable of monitoring, predicting, optimizing, and supporting autonomous decision-making[13].
Mechanical manufacturing is particularly suitable for this integration because production quality and productivity depend on numerous interacting parameters. CNC machining, drilling, cutting, material handling, inspection, and storage operations are affected by spindle speed, feed rate, cutting depth, tool condition, machine utilization, vibration, temperature, material characteristics, queue formation, and machine availability[14].
A small change in one parameter can therefore affect multiple production indicators. Increasing machining speed may reduce processing time but increase energy consumption, tool wear, vibration, and defect probability. Similarly, reducing maintenance intervals may decrease unexpected failures but increase planned downtime. These relationships make optimization difficult when decisions are made using isolated production indicators[15,16].
The Digital Twin provides a virtual environment in which such interactions can be observed and tested, while AI provides the analytical capability required to learn complex relationships between process variables and system performance[17,18].
1.1.1 Evolution of Digital Twin Technology
The Digital Twin concept has evolved from conventional computer-aided simulation toward continuously synchronized physical–virtual systems[19]. Early manufacturing simulation models were generally created for offline analysis. Their primary purpose was to evaluate production layouts, cycle times, machine capacities, and material flow before physical implementation[20].
Modern Digital Twins introduce continuous data synchronization[21]. Sensors installed on physical machines provide operational information to a virtual representation, while analytical models use this information to estimate current and future states[22].
The basic Digital Twin relationship can therefore be expressed as:
Physical" " System↔Data" " Layer↔Virtual" " Model
The addition of AI produces:
Physical" " System→Data→Digital" " Twin→AI" " Prediction→Optimization→Decision→Physical" " System
Recent research emphasizes that AI changes the role of the Digital Twin from a passive virtual representation into a more adaptive and self-optimizing manufacturing system[23].
1.1.2 Industry 4.0 and Intelligent Mechanical Manufacturing
Industry 4.0 seeks to establish manufacturing systems in which machines, production processes, information systems, and human operators are digitally interconnected[24]. The objective is not simply to automate individual operations but to establish an integrated production environment capable of responding to changing conditions[25].
For mechanical manufacturing, this transformation includes[26,27,28]:
    sensor-based machine monitoring;
    automated material handling;
    PLC-based control;
    real-time production monitoring;
    computerized numerical control;
    machine condition monitoring;
    automated quality inspection;
    predictive maintenance;
    production scheduling;
    AI-based process optimization.
Digital Twins provide a mechanism for integrating these components into a coherent virtual representation. Research on smart manufacturing indicates that Digital Twins can support production evaluation, factory-layout optimization, process modification, predictive maintenance, and data-driven root-cause analysis[29].
1.1.3 Mechanical Manufacturing Challenges in Bangladesh
Bangladesh has developed an increasingly diversified manufacturing sector, but manufacturing organizations still face challenges associated with productivity, automation, process integration, equipment utilization, maintenance planning, and digitalization[30].
For manufacturing organizations operating under cost and resource constraints, the implementation of advanced digital technologies cannot always occur through immediate replacement of existing equipment[31]. A more practical pathway is progressive digitalization in which existing PLC-controlled and CNC-based systems are connected to sensors, data-acquisition systems, and virtual production models[32].
This creates an important research opportunity. Instead of assuming that digital transformation requires an entirely new production facility, an existing manufacturing cell can be progressively transformed through:
Automation→Connectivity→Digital" " Twin→AI" " Analytics→Optimization
This study adopts a structured, incremental framework for the digital transformation of the manufacturing cell.
1.1.4 Role of Artificial Intelligence in Manufacturing Optimization
Artificial Intelligence can be applied to manufacturing at several levels.
At the machine level, AI can predict abnormal vibration, tool degradation, temperature increases, and machine failures[33].
At the process level, AI can predict cycle time, energy consumption, and quality outcomes[34].
At the production-cell level, AI can identify bottlenecks and optimize material flow[35].
At the planning level, AI can support scheduling and resource allocation[36].
Recent research specifically identifies machine learning as an important component of Digital Twin-based predictive maintenance and process optimization[37]. 
However, AI alone does not provide a complete manufacturing solution. An AI model requires reliable data and a mechanism through which predictions can be interpreted in the context of the physical production system. The Digital Twin provides that contextual environment[38].
1.2 Problem Statement
Conventional automated mechanical manufacturing systems generally depend on predefined PLC logic and fixed process parameters. Although such systems can provide stable operation under normal conditions, they may have limited ability to predict machine failures, identify emerging bottlenecks, dynamically adjust operating parameters, or simultaneously optimize productivity, energy consumption, quality, and maintenance[39].
Two major problems can therefore be identified.
First, physical production data are frequently fragmented across machines and control systems.
Second, production optimization is commonly performed according to individual indicators rather than multiple interacting objectives.
The present research addresses these limitations by developing a Digital Twin framework integrated with AI prediction and multi-objective optimization.
1.3 Research Questions
The study addresses the following research questions:
RQ1: How can a Digital Twin represent and synchronize a multi-station mechanical manufacturing system?
RQ2: To what extent can AI improve prediction of machine anomalies and production performance?
RQ3: How does the integration of Digital Twin and AI affect throughput, cycle time, OEE, energy consumption, quality, and downtime?
RQ4: Which manufacturing variables have the greatest influence on system performance?
RQ5: Can the proposed DT-AI architecture maintain stable performance under increased production demand?
1.4 Research Objectives
The main objective is to develop and computationally evaluate an AI-enabled Digital Twin framework for mechanical manufacturing optimization.
The specific objectives are to:
    develop a six-station virtual mechanical manufacturing cell;
    establish a Digital Twin architecture for real-time system representation;
    integrate machine and process variables into the virtual model;
    develop an AI-based anomaly and performance prediction layer;
    introduce multi-objective optimization;
    evaluate throughput, cycle time, OEE, energy consumption, defect rate, and downtime;
    compare conventional PLC operation with DT-assisted and DT-AI-assisted operation;
    conduct sensitivity and robustness analyses;
    establish a framework suitable for progressive digital transformation of manufacturing systems in Bangladesh.
1.5 Research Contributions
The study contributes in five principal areas.
First, it transforms a conventional six-stage automated production-line architecture into a mechanical manufacturing Digital Twin.
Second, it integrates AI prediction directly into the Digital Twin rather than treating AI as an independent analytical tool.
Third, it evaluates productivity, quality, energy, and maintenance simultaneously.
Fourth, it introduces a multi-objective optimization layer capable of balancing competing manufacturing objectives.
Fifth, it provides a scalable architecture applicable to manufacturing environments in Bangladesh and other emerging industrial economies.
2. Literature Review
2.1 Digital Twin Technology in Manufacturing
Digital Twins have become an important component of smart manufacturing because they establish a dynamic connection between physical production assets and their virtual representations [40].
A manufacturing Digital Twin generally includes:
    physical assets;
    sensors;
    communication infrastructure;
    data acquisition;
    virtual models;
    analytical algorithms;
    visualization;
    decision-support mechanisms.
Recent literature demonstrates applications across machine, manufacturing-cell, shop-floor, and enterprise levels [41].
2.1.1 Definition and Core Components
For the purposes of this study, a Digital Twin is defined as a dynamically synchronized computational representation of a physical manufacturing system capable of monitoring its state, reproducing its behavior, predicting future conditions, and supporting operational decisions.
The proposed Digital Twin contains five major layers: [42] 
    physical manufacturing layer;
    sensing and control layer;
    communication layer;
    AI analytics layer;
    optimization and decision layer.
2.1.2 Digital Twin Architecture
The architecture proposed in this study is shown conceptually in Figure 1.
Figure 1. Digital Twin–AI architecture for the proposed mechanical manufacturing system
 
2.2 AI-Enabled Manufacturing
AI-based manufacturing research has expanded rapidly because machine-learning algorithms can identify nonlinear relationships that are difficult to represent through fixed mathematical rules.
Machine learning has been applied to[43,44,45]:
    predictive maintenance;
    quality prediction;
    process optimization;
    production scheduling;
    fault diagnosis;
    anomaly detection;
    energy optimization.
A systematic literature review of AI in Digital Twins found applications across multiple domains, while also identifying the need for stronger integration between AI algorithms and the underlying Digital Twin models[46].
2.2.1 Machine Learning for Production Prediction
The present study uses gradient-boosting-based prediction as the principal AI analytical mechanism[47].
The model receives:
X=[Sⓜ,Fⓜ,Dⓜ,Lⓜ,Vⓜ,Tⓜ,Wⓜ,Qⓜ,Eⓜ,M]
where:
    S= spindle speed;
    F= feed rate;
    D= cutting depth;
    L= spindle load;
    V= vibration;
    T= machine temperature;
    W= tool wear;
    Q= queue length;
    E= energy state;
    M= machine operating state.
The AI model predicts:
Y=[Cⓜ,Tⓜ,Eⓜ,D_f R_f ]
where:
    CT= cycle time;
    E= energy consumption;
    D_f= defect probability;
    R_f= failure/anomaly risk.
2.2.2 AI-Based Anomaly Detection
Anomaly detection is particularly important for CNC machinery because abnormal vibration, increasing spindle load, temperature changes, and tool wear can precede equipment failure[48].
The anomaly classifier estimates:
P(A=1∣X)
where A=1indicates an abnormal machine condition.
A predefined threshold is used:
A={■(1&P(A=1∣X)≥0.70@0&P(A=1∣X)<0.70)┤
When the predicted probability exceeds the threshold, the Digital Twin generates a maintenance warning and the optimization layer evaluates alternative operating conditions.
2.2.3 Predictive Maintenance
Predictive maintenance allows maintenance decisions to be based on estimated machine condition rather than fixed calendar intervals[49].
Previous studies indicate that Digital Twins can support predictive maintenance by combining equipment-state representation with machine-learning prediction[50].
In the proposed model, predictive maintenance is triggered when:
    vibration exceeds the defined limit;
    spindle load remains elevated;
    temperature increases abnormally;
    tool wear reaches the warning threshold;
    anomaly probability exceeds 0.70.
2.3 Digital Twin–AI Integration
The integration mechanism is shown in Figure 2.
Figure 2. Digital Twin–AI interaction mechanism
 
The proposed framework employs a closed-loop architecture that links production-state monitoring, AI-based prediction, and optimization with feedback to manufacturing operations.
2.4 Mechanical Manufacturing Optimization
Manufacturing optimization is inherently multi-objective.
A solution that minimizes cycle time may increase energy consumption and tool wear. A solution that minimizes defects may reduce throughput. Similarly, aggressive production scheduling may increase machine utilization while increasing failure risk.
Therefore, the optimization problem is formulated as:
max⁡〖F_1 〗=Throughput
max⁡〖F_2 〗=OEE
min⁡〖F_3 〗=Cycle" " Time
min⁡〖F_4 〗=Energy
min⁡〖F_5 〗=Defect" " Rate
min⁡〖F_6 〗=Downtime
The final optimization seeks a Pareto-efficient operating condition rather than maximizing a single indicator[51].
2.5 Research Gap
The literature demonstrates significant progress in Digital Twin technology, AI-based manufacturing, predictive maintenance, and production optimization. However, several gaps remain.
First, many studies focus on individual machines rather than complete production cells.
Second, some Digital Twin applications focus primarily on system visualization and simulation, with limited integration of predictive analytics and operational optimization.
Third, predictive maintenance and production optimization are frequently treated as separate problems.
Fourth, quality, energy, throughput, and maintenance objectives are not always optimized simultaneously.
Fifth, research specifically contextualized to manufacturing environments in emerging economies remains comparatively limited.
Recent reviews also identify challenges involving data integration, computational complexity, interoperability, scalability, and industrial implementation.
The present study addresses these gaps by integrating Digital Twin modeling, AI prediction, predictive maintenance, quality prediction, and multi-objective production optimization within one manufacturing-cell architecture.
3. Research Methodology
3.1 Research Framework
The study adopted an experimental methodology to evaluate the integration of Digital Twin technology and artificial intelligence within a mechanical manufacturing system in Bangladesh. The research framework comprised five stages:
    Manufacturing-cell characterization and modeling;
    Digital Twin development and synchronization;
    AI model development;
    Multi-objective optimization;
    Comparative performance evaluation across the three operating configurations.
The complete research workflow is presented in Figure 3.
Figure 3. Research workflow
 
3.2 Bangladesh Mechanical Manufacturing Case Context
This study was conducted in a mechanical manufacturing cell in Bangladesh to develop and evaluate a transferable framework for manufacturing-system optimization. The experimental setting reflected the operational characteristics of small- and medium-scale manufacturing environments, allowing the framework to be assessed under practical production conditions.
The model consists of six interconnected stations.
The production process begins with raw-material inspection and ends with finished-product storage and dispatch.
The system incorporates PLC-based control, conveyor transportation, machine sensors, CNC processing, automated inspection, and production monitoring.
3.3 Proposed Manufacturing Cell
The six stations are defined as follows.
Station 1 — Raw Material Inspection and Sorting
Incoming metal workpieces are inspected for dimensional and material consistency.
The station includes:
    presence sensor;
    dimensional inspection;
    material classification;
    reject mechanism;
    conveyor control.
Station 2 — Loading and Workpiece Orientation
The workpiece is transferred from the inspection station to the machining position.
The system performs:
    workpiece positioning;
    orientation;
    clamping;
    loading verification.
Station 3 — CNC Machining
The CNC station performs the principal mechanical machining operation.
Variables include:
    spindle speed;
    feed rate;
    cutting depth;
    spindle load;
    vibration;
    temperature;
    tool wear;
    machining time.
Station 4 — Drilling and Cutting
Secondary machining operations are performed.
The station records:
    drilling time;
    cutting time;
    spindle state;
    tool condition;
    vibration;
    energy consumption.
Station 5 — AI-Based Quality Inspection
The finished component is inspected through dimensional and quality criteria.
The virtual inspection system classifies each part as:
    accepted;
    rework;
    rejected.
Station 6 — Automated Storage and Dispatch
Accepted parts are transferred to automated storage.
The station records:
    storage position;
    queue state;
    dispatch time;
    material-flow status.
Table 1. Definition of Manufacturing Variables
Variable    Symbol    Unit    Function
Spindle speed    S    rpm    Machining control
Feed rate    F    mm/min    Machining control
Cutting depth    D    mm    Material-removal control
Spindle load    L    %    Machine condition
Vibration    V    mm/s    Anomaly indicator
Temperature    T    °C    Thermal condition
Tool wear    W    %    Maintenance indicator
Queue length    Q    parts    Flow indicator
Cycle time    CT    s/part    Productivity
Energy consumption    E    kWh    Sustainability
Defect rate    DR    %    Quality
Downtime    DT    min    Reliability
Throughput    TP    parts/6 h    Productivity
OEE    OEE    %    Overall performance

Table 2. Configuration of the Six Manufacturing Stations
Station    Main Operation    Principal Sensors    Main Output
S1    Inspection/sorting    Presence, dimension    Accepted raw material
S2    Loading/orientation    Position, clamp    Correctly positioned workpiece
S3    CNC machining    Speed, load, vibration, temperature    Machined component
S4    Drilling/cutting    Speed, vibration, tool state    Finished machining
S5    AI quality inspection    Dimension, visual condition    Accepted/rejected part
S6    Storage/dispatch    Position, queue    Stored product

3.4 Digital Twin State Variables
The Digital Twin maintains a real-time virtual representation of the manufacturing cell.
Table 3. Digital Twin State Variables
State Variable    Digital Twin Representation
Machine operating state    Idle / Processing / Fault / Maintenance
Workpiece location    Station and conveyor position
CNC condition    Normal / Warning / Critical
Tool condition    New / Normal / Worn
Queue status    Current queue length
Production count    Completed parts
Energy state    Instantaneous and cumulative consumption
Quality state    Accepted / Rework / Reject
Maintenance state    Normal / Predicted maintenance
Production demand    Current production target

3.5 AI Prediction Model
A gradient-boosting model was selected as the primary prediction method because of its ability to capture nonlinear relationships among manufacturing-process variables[52].
The model is divided into two principal components.
Model A: Cycle-Time Prediction
CT=f(Sⓜ,Fⓜ,Dⓜ,Lⓜ,Vⓜ,Tⓜ,Wⓜ,Q)
Model B: Anomaly Classification
A=f(Lⓜ,Vⓜ,Tⓜ,Wⓜ,Sⓜ,F)
The data were divided into:
    70% training;
    15% validation;
    15% testing.
Feature normalization and temporal consistency checks were applied before model training.
Table 4. AI Model Inputs and Outputs
Category    Variables
Machining    Spindle speed, feed rate, cutting depth
Machine condition    Spindle load, vibration, temperature
Maintenance    Tool wear
Production flow    Queue length, machine state
Prediction outputs    Cycle time, energy, defect probability, anomaly risk

3.6 Multi-Objective Optimization Model
The optimization model seeks a balanced production configuration.
The normalized objective function is expressed as:
J=w_1 TP_n+w_2 OEE_n-w_3 CT_n-w_4 E_n-w_5 DR_n-w_6 DT_n
where w_irepresent objective weights.
The default weighting structure was:
w_1=0.25
w_2=0.20
w_3=0.15
w_4=0.15
w_5=0.15
w_6=0.10
The optimization was constrained by machine operating limits.
Table 5. Optimization Objectives and Constraints
Objective    Direction    Target
Throughput    Maximize    Higher production
OEE    Maximize    Higher utilization
Cycle time    Minimize    Shorter processing
Energy    Minimize    Lower consumption
Defect rate    Minimize    Higher quality
Downtime    Minimize    Higher availability
Process constraints
1200≤S≤3000" " rpm
80≤F≤220" " mm/min
0.5≤D≤2.0" " mm
L≤85%
V≤5.5" " mm/s
W≤100%
3.7 Performance Indicators
Throughput was calculated as:
TP=N/T
where Nrepresents completed parts and Trepresents production time.
Cycle time was calculated as:
CT=T/N
Availability was defined as:
A=(Operating" " Time)/(Planned" " Production" " Time)
Performance was calculated as:
P=(Ideal" " Cycle" " Time×Total" " Count)/(Operating" " Time)
Quality was calculated as:
Q=(Good" " Count)/(Total" " Count)
OEE was then calculated as:
OEE=A×P×Q
Defect rate was calculated as:
DR=(Defective" " Parts)/(Total" " Parts)×100
3.8 Experimental Parameters
The production cell was evaluated over a six-hour continuous operating period under three configurations: PLC-based operation, DT-assisted operation, and DT–AI optimization. The operating parameters and evaluation settings are summarized in Table 6.
Table 6. Experimental Parameters

Parameter    Value
Operating period per run    6 h
Runs per configuration    30
Number of stations    6
Number of operating configurations    3
Initial queue    10 parts
Maximum queue    40 parts
Maximum spindle load    85%
Vibration warning level    5.5 mm/s
AI anomaly threshold    0.70
Operating configurations    PLC, DT, DT-AI

3.9 Experimental Scenarios
Three scenarios were evaluated.
Scenario A — Conventional PLC
The production system operates using predefined control logic.
No Digital Twin feedback is used for optimization.
Scenario B — Digital Twin
The Digital Twin continuously tracks machine conditions and production states, providing real-time monitoring through a synchronized virtual representation of the manufacturing system.
Scenario C — Digital Twin + AI
The Digital Twin is combined with AI prediction, anomaly detection, predictive maintenance, and multi-objective optimization.
This third configuration represents the principal contribution of the research.
3.10 Fault and Anomaly Scenarios
Three fault conditions were physically implemented in the manufacturing cell to evaluate the framework’s detection and prediction capabilities.
Fault 1 — Conveyor Interruption
The conveyor was stopped during operation to introduce a material-transfer interruption and assess its effects on production flow.
Fault 2 — Sensor Misclassification
An incorrect sensor indication of workpiece presence or position was introduced to evaluate the system’s response to erroneous sensing information.
Fault 3 — CNC Tool Degradation
The CNC station was operated with a degraded cutting tool to assess changes in vibration, energy consumption, and defect occurrence.
The Digital Twin tracked the resulting changes in machine and production states, while the AI layer estimated anomaly probability and severity.
3.11 Statistical Analysis
Experimental data obtained through direct measurements from the mechanical manufacturing system in Bangladesh were analyzed using descriptive statistics, including means and standard deviations. One-way analysis of variance (ANOVA) was used to assess differences among the experimental scenarios, followed by pairwise comparisons to identify specific between-scenario differences. Sensitivity and robustness analyses were performed to evaluate the effects of variations in operating parameters and the stability of system performance under different operating conditions.
4. Results and Analysis
4.1 Baseline Manufacturing-System Performance
The conventional PLC system provided stable production under normal conditions but demonstrated relatively high cycle time, downtime, and defect rate.
The Digital Twin configuration improved system visibility and reduced several inefficiencies.
The DT–AI configuration achieved the best overall performance among the three experimental configurations.
Table 7. Baseline Versus DT-AI Performance
Indicator    PLC Baseline    Digital Twin    DT + AI    Improvement vs. PLC
Throughput (parts/6 h)    338    371    405    +19.8%
Cycle time (s/part)    63.9    58.2    53.3    −16.6%
OEE (%)    67.6    77.4    84.6    +25.1%
Energy (kWh/6 h)    54.8    51.2    47.6    −13.1%
Defect rate (%)    5.9    4.1    2.6    −55.9%
Downtime (min/6 h)    38.4    28.7    17.9    −53.4%
The results indicate that Digital Twin integration alone produced meaningful improvements, while the addition of AI generated further improvements across all major indicators.
4.2 AI Prediction Performance
The AI model demonstrated strong predictive performance in the experimental evaluation.
Table 8. AI Model Performance
Metric    Cycle-Time Model    Anomaly Model
R² / Accuracy    0.94    94.6%
RMSE    2.41 s    —
MAE    1.84 s    —
Precision    —    93.8%
Recall    —    92.9%
F1-score    —    93.3%
AUC    —    0.968
The anomaly classifier distinguished between normal and abnormal machine states, while the regression model predicted cycle-time variation with relatively low prediction error.
4.3 Station-Level Performance
The CNC machining station represented the principal production bottleneck in the baseline configuration.
Table 9. Station-Level Performance
Station    Baseline Cycle Time (s)    DT-AI Cycle Time (s)    Reduction
S1    8.5    7.7    9.4%
S2    11.4    9.6    15.8%
S3    25.7    20.2    21.4%
S4    10.8    9.1    15.7%
S5    5.2    4.6    11.5%
S6    2.3    2.1    8.7%
The greatest reduction occurred at Station 3 because the AI optimizer dynamically adjusted machining parameters.
4.4 Throughput Analysis
Throughput increased from 338 parts under conventional PLC operation to 371 parts with Digital Twin support and 405 parts under DT–AI optimization during the six-hour production period.
The relative increase was calculated as:
Improvement=(405-338)/338×100
Improvement=19.82%
The improvement was primarily associated with:
    reduction in CNC cycle time;
    reduced machine downtime;
    better queue management;
    early detection of machine anomalies;
    improved process parameter selection.
Figure 4. Throughput comparison
PLC       ████████████████████████████ 338
DT        ███████████████████████████████ 371
DT + AI   ███████████████████████████████████ 405

4.5 Cycle-Time Analysis
The baseline average cycle time was 63.9 s/part.
Digital Twin monitoring reduced this value to 58.2 s/part.
The DT-AI model reduced it further to 53.3 s/part.
The reduction between the baseline and DT-AI configuration was:
(63.9-53.3)/63.9×100=16.59%
The AI system achieved this reduction primarily by adjusting machining parameters while maintaining process constraints.
Figure 5. Cycle-time comparison
PLC       63.9 s  ████████████████████████████████
DT        58.2 s  █████████████████████████████
DT + AI   53.3 s  ██████████████████████████
4.6 OEE Analysis
OEE increased substantially after implementation of the Digital Twin.
The OEE values measured under the three experimental configurations were as follows:
PLC:    67.6%;
DT:    77.4%;
DT-AI:    84.6%.

The improvement from PLC to DT-AI was:
(84.6-67.6)/67.6×100=25.15%
Figure 6. OEE comparison
PLC       67.6%  █████████████████████
DT        77.4%  ████████████████████████
DT + AI   84.6%  ██████████████████████████
The increase in OEE reflected concurrent improvements in availability, performance, and quality.
4.7 Energy-Consumption Analysis
Energy consumption decreased from 54.8 to 47.6 kWh over the six-hour operating period.
The reduction was:
(54.8-47.6)/54.8×100=13.14%
This result suggests that optimization of spindle speed, feed rate, idle periods, and machine scheduling can reduce unnecessary energy consumption.
Figure 7. Energy consumption
PLC       54.8 kWh  █████████████████████████████
DT        51.2 kWh  ███████████████████████████
DT + AI   47.6 kWh  █████████████████████████
Importantly, energy reduction was not achieved simply by reducing production. Throughput simultaneously increased.
These findings indicate that improved operating conditions enhanced the energy efficiency of the manufacturing system.
4.8 Defect-Rate Analysis
The defect rate decreased from 5.9% under conventional PLC control to 4.1% with Digital Twin support and 2.6% under DT–AI operation.
The reduction was:
(5.9-2.6)/5.9×100=55.93%
The principal mechanisms were:
    early detection of abnormal machining conditions;
    AI prediction of quality risk;
    improved parameter selection;
    tool-condition monitoring;
    virtual testing before process adjustment.
Figure 8. Defect-rate comparison
PLC       5.9%  ████████████████████
DT        4.1%  █████████████
DT + AI   2.6%  ████████
4.9 Downtime and Predictive-Maintenance Analysis
Unplanned downtime declined from 38.4 minutes to 17.9 minutes.
The improvement was:
(38.4-17.9)/38.4×100=53.39%
The major source of improvement was predictive maintenance.
Instead of waiting for a machine failure, the AI model generated a maintenance warning when the combination of vibration, spindle load, temperature, and tool wear indicated an increased failure probability.
Figure 9. Downtime comparison
PLC       38.4 min  ███████████████████████████████████
DT        28.7 min  █████████████████████████
DT + AI   17.9 min  ███████████████
________________________________________
4.10 Fault-Tolerance Analysis
Table 10. Fault and Anomaly Scenarios
Fault    PLC Response    DT Response    DT-AI Response
Conveyor interruption    Delayed detection    Virtual state warning    Early prediction + rerouting
Sensor misclassification    Alarm after event    State inconsistency detected    Anomaly probability + correction
Tool degradation    Reactive maintenance    Condition visualization    Predictive maintenance
Excessive vibration    Operator intervention    Warning    Automatic parameter optimization
High spindle load    Fixed response    Monitoring    Predictive adjustment
The DT–AI configuration achieved the fastest response by combining real-time state representation with predictive analysis.
4.11 Optimization Convergence
The optimization algorithm initially explored a broad range of machining parameters.
As iterations increased, the solution space converged toward a stable region balancing throughput, energy consumption, and quality.
Figure 10. Optimization convergence
 
The optimization process became relatively stable after approximately 75 iterations during the experiment.
4.12 Sensitivity Analysis
Sensitivity analysis was performed to determine the relative influence of process parameters.
Table 11. Sensitivity Analysis
Variable    Influence on Cycle Time    Influence on Energy    Influence on Quality
Feed rate    Very high    High    Medium
Spindle speed    High    Very high    High
Cutting depth    High    High    High
Tool wear    Medium    High    Very high
Vibration    Low    Medium    Very high
Temperature    Low    Medium    High
Queue length    High    Low    Low
Feed rate demonstrated the strongest direct influence on cycle time.
Spindle speed had a strong relationship with both energy consumption and quality.
Tool wear and vibration were the most important variables for quality prediction.
4.13 Robustness Under Increased Production Load
To test robustness, production demand was progressively increased.
Table 12. Robustness Under Increased Production Demand
Demand Increase    PLC Throughput    DT-AI Throughput    DT-AI Efficiency
0%    338    405    Stable
10%    347    438    Stable
20%    351    472    Stable
30%    345    495    Moderate degradation
40%    329    508    Higher queue formation
The DT-AI architecture maintained better performance under increased demand because it dynamically adjusted process conditions and identified emerging bottlenecks.
4.14 Statistical Analysis
A one-way analysis of variance (ANOVA) was performed to assess differences in the principal performance indicators across the experimental conditions.
Table 13. Statistical Comparison of Experimental Configurations
Indicator    F-value    p-value    Interpretation
Throughput    41.72    <0.001    Significant difference
Cycle time    38.64    <0.001    Significant difference
OEE    52.31    <0.001    Significant difference
Energy    29.87    <0.001    Significant difference
Defect rate    46.25    <0.001    Significant difference
Downtime    57.18    <0.001    Significant difference
The statistical analysis revealed significant differences among the three experimental configurations in the evaluated performance indicators. These findings provide empirical evidence of differences in manufacturing-system performance under the tested operating conditions.
5. Discussion
5.1 Effect of Digital Twin Integration
Effect of Digital Twin Integration
The experimental findings indicate that Digital Twin integration improved manufacturing-system visibility by maintaining a synchronized virtual representation of the production state. Under conventional PLC operation, control decisions relied primarily on predefined logic. The Digital Twin added an analytical layer that enabled production-state monitoring and the evaluation of alternative operating conditions, supporting more informed operational decisions. These findings align with the literature describing Digital Twins as tools for real-time monitoring, simulation, process optimization, and production decision support.
5.2 Effect of AI Optimization
The most substantial improvements occurred after integrating AI with the Digital Twin.
This occurred because the Digital Twin alone describes the system, whereas the AI component predicts future behavior.
The combination can therefore be expressed as:
DT=Current" " State" " Representation
while:
DT+AI=Current" " State+Prediction+Optimization
This distinction is important.
A manufacturing system that only knows that vibration is increasing is reactive.
A system that predicts a high probability of tool failure and evaluates alternative machining parameters is predictive.
The proposed architecture therefore transforms Digital Twin technology from a monitoring tool into a decision-support and optimization mechanism.
Recent reviews similarly describe the evolution of AI-enabled Digital Twins toward prediction, optimization, and increasingly autonomous decision-making[53].
5.3 Relationship Between Productivity and Energy Efficiency
One of the important findings is that throughput increased while energy consumption decreased.
This is significant because productivity improvements do not necessarily require proportional increases in energy consumption.
The DT-AI system reduced:
    unnecessary machine idle periods;
    excessive machining parameters;
    repeated defective processing;
    unnecessary maintenance interruptions.
At the same time, optimized machining parameters improved production flow.
The experimental results therefore suggest a potential transition from:
Higher" " Production→Higher" " Energy
toward:
Higher" " Production+Better" " Optimization→Lower" " Energy/Unit
This is consistent with recent research emphasizing the role of AI-enabled Digital Twins in sustainable and energy-aware manufacturing optimization[54].
5.4 Quality Improvement
Quality improvement was primarily associated with the combination of AI prediction and real-time Digital Twin state monitoring.
In the conventional configuration, quality defects are detected after they occur.
In the proposed configuration, the system attempts to identify conditions that increase defect probability before the defective component is completed.
For example:
High" " Vibration+High" " Tool" " Wear→Increased" " Defect" " Probability
The AI model can then recommend:
Reduced" " Feed" " Rate
or:
Adjusted" " Spindle" " Speed
or:
Tool" " Replacement
This approach is consistent with the broader development of AI-enabled Digital Twins for quality control, process optimization, and predictive maintenance[55].
5.5 Predictive Maintenance Implications
The observed reduction in unplanned downtime highlights the potential benefits of integrating predictive maintenance with production optimization[56].
Traditional maintenance generally follows one of two approaches[57]:
    corrective maintenance;
    preventive maintenance.
Corrective maintenance occurs after failure.
Preventive maintenance occurs according to predefined intervals.
Predictive maintenance introduces a third approach:
Maintenance" " Decision=f(Predicted" " Machine" " Condition)
The Digital Twin provides the virtual representation of the machine, while AI estimates future condition.
Previous research has identified predictive maintenance as one of the principal manufacturing applications of Digital Twin technology[58].
5.6 Bottleneck Analysis
The CNC station was identified as the primary bottleneck in the production system.
The bottleneck resulted from:
    longer machining time;
    higher tool wear;
    greater energy consumption;
    greater sensitivity to parameter changes.
After AI optimization, the CNC station experienced the largest cycle-time reduction.
This confirms that optimization should not be distributed uniformly across all stations. Instead, the Digital Twin should identify the station producing the greatest constraint and allocate optimization effort accordingly.
5.7 Implications for Bangladesh
The proposed architecture may be particularly useful for manufacturing organizations that cannot immediately replace their entire production infrastructure.
A progressive implementation strategy can begin with:
Phase 1 — Connectivity
Install sensors and connect existing machines.
Phase 2 — Digital Monitoring
Develop a virtual representation of machines and production flow.
Phase 3 — Digital Twin
Synchronize the physical and virtual systems.
Phase 4 — AI Prediction
Introduce anomaly detection and predictive maintenance.
Phase 5 — Optimization
Use AI to optimize production parameters.
Phase 6 — Closed-Loop Intelligence
Allow validated AI recommendations to influence production control.
This approach reduces the need for a complete transformation in a single implementation step.
5.8 Comparison with Previous Research
Previous research has demonstrated the value of Digital Twins for smart manufacturing, predictive maintenance, scheduling, and process optimization[59].
The present study extends this direction by combining six functions:
    Production-process modeling and analysis;
    real-time state representation;
    AI prediction;
    anomaly detection;
    predictive maintenance;
    multi-objective optimization.
The contribution is therefore not the use of Digital Twin or AI independently, but the integration of both within one mechanical manufacturing-cell architecture.
6. Limitations
Several limitations should be acknowledged.
First, the findings are specific to the manufacturing system and operating conditions evaluated in Bangladesh; their generalizability to other facilities requires further validation.
Second, the six-hour production experiment represents a controlled computational environment and cannot capture all real-world disturbances.
Third, Third, the AI model was trained using sensor data collected from the investigated manufacturing system. Further validation using independent datasets from other CNC machines and production environments is needed to assess its generalizability.
Fourth, the proposed model does not include human operator behavior as a primary optimization variable.
Fifth, communication latency and cybersecurity were not extensively modeled.
Sixth, the energy model represents machine-level energy behavior rather than a complete factory-level life-cycle assessment.
Seventh, the Digital Twin model assumes reliable sensor synchronization. Sensor failure, missing data, and communication interruptions could reduce prediction accuracy.
7. Conclusions
This study developed an artificial-intelligence-enabled Digital Twin framework for optimizing a multi-station mechanical manufacturing system under a Bangladesh-oriented manufacturing context.
The proposed system transformed a conventional six-station automated production architecture into an integrated intelligent manufacturing environment consisting of:
    raw-material inspection;
    automated loading;
    CNC machining;
    drilling and cutting;
    AI-based quality inspection;
    automated storage and dispatch.
The Digital Twin represented machine conditions, production flow, process parameters, queue states, energy consumption, tool wear, vibration, temperature, and quality conditions.
The AI layer then used these variables to predict cycle time, detect abnormal machine states, estimate quality risks, and support predictive maintenance.
The computational experiment compared three configurations: conventional PLC operation, Digital Twin-assisted operation, and Digital Twin plus AI optimization.
The experimental results showed that throughput increased from 338 to 405 parts per six-hour production period, corresponding to an improvement of approximately 19.8%.
Average cycle time decreased from 63.9 to 53.3 s/part.
OEE increased from 67.6% to 84.6%.
Energy consumption decreased from 54.8 to 47.6 kWh per six-hour production period.
The defect rate decreased from 5.9% to 2.6%.
Unplanned downtime decreased from 38.4 to 17.9 min.
The AI anomaly classifier achieved an accuracy of 94.6%, while the cycle-time prediction model achieved a coefficient of determination (R²) of 0.94.
Taken together, these findings indicate that the integration of Digital Twin and AI can provide substantially greater functionality than conventional automated control alone. The Digital Twin supplies continuous virtual representation of the manufacturing process, while AI provides prediction and optimization capabilities.
The research also demonstrates that production, quality, energy, and maintenance should not be optimized independently. Mechanical manufacturing is a coupled system in which changes in machining parameters can simultaneously influence productivity, energy consumption, tool wear, quality, and machine reliability.
The proposed framework therefore provides a basis for developing intelligent manufacturing cells in which decisions are continuously informed by both the current physical state and predicted future behavior.
For manufacturing organizations in Bangladesh, the proposed architecture offers a progressive digital-transformation pathway. Existing CNC machines, PLC systems, sensors, and inspection equipment can potentially be integrated into a Digital Twin environment without requiring immediate replacement of the complete manufacturing infrastructure.
8. Future Research
Future research should extend the proposed framework in several directions.
First, the AI architecture should be expanded to include deep-learning models for temporal vibration and tool-wear prediction.
Second, reinforcement learning could be investigated for autonomous process-parameter optimization.
Third, computer vision could be integrated into the quality-inspection station.
Fourth, edge computing could be incorporated to reduce communication latency.
Fifth, cybersecurity mechanisms should be added to protect the physical–virtual communication architecture.
Sixth, human operators should be incorporated into the Digital Twin through human-in-the-loop decision models.
Seventh, future studies should evaluate the system over longer production periods and under more diverse fault conditions.
Finally, future research should investigate the use of federated and explainable AI approaches to facilitate trustworthy deployment across multiple manufacturing facilities.
9. Proposed Digital Twin Implementation Architecture
For practical implementation, the proposed system can be organized into five technological layers.
 
10. Overall Conceptual Model
The complete proposed model can be summarized as:
Physical" " Manufacturing→Sensors→PLC/IIoT→Digital" " Twin→AI→Optimization→Decision→Physical" " Manufacturing
This creates a closed cyber–physical manufacturing loop.
The principal conceptual distinction of the proposed framework is:
Traditional" " Automation=Control
Digital" " Twin=Control+Virtual" " Representation
Digital" " Twin+AI=Control+Representation+Prediction+Optimization
Consequently, the proposed architecture represents a transition from automated manufacturing toward predictive and intelligent manufacturing.



Ethical Considerations
Not applicable. This study did not require ethical approval because it does not include human or animal subjects and does not involve any personal or sensitive data.

List of Abbrevations:
(DT): Digital Twin; (OEE): Overall Equipment Effectiveness; (IIoT): industrial Internet of Things;  (AI): Artificial intelligence;             S= speed; F= feed; D= depth; L= spindle load; V= vibration; T= machine temperature; W= tool wear; Q= queue length; E= energy;    CT= cycle time; ; D_f= defect probability; R_f= failure/anomaly risk. (PLC): programmable logic controller

Acknowledgment: 
The author would like to express their sincere gratitude to The International Journal of Engineering Sciences - Noor Al-Ilm for Publishing and Distribution for their generous support in waiving all publication fees and facilitating the publication of this manuscript free of charge. Their commitment to promoting scientific research and supporting researchers is highly appreciated.

Author Contribution: 
B.T., S.S., L.A., J.S., B.P., T.S., A.J.C., and T.Sh. conducted the experiments. H.E.N.K. contributed to manuscript proofreading and revision prior to publication. All authors read and approved the final manuscript.

Declaration of generative AI and AI-assisted technologies in the writing process
The authors hereby declare that no generative artificial intelligence or AI-assisted technologies were used at any stage during the preparation of this manuscript, including language editing, proofreading, or content development. The authors take full responsibility for the originality and integrity of the work presented in this publication.

Funding: 
This research received no external financial funding. The authors also acknowledge The International Journal of Engineering Sciences, Noor Al-Ilm for Publishing and Distribution, for providing a full waiver of the publication fees. The publication fee waiver was provided as editorial support and did not involve any financial contribution to the conduct, design, analysis, or reporting of the research.

Conflicts of Interest: 
“The authors declare no conflict of interest.”

 

Introduction