11. Simulation, Analytics and Facility Decision-Making

Introduction and Context

Facility planning decisions are often made before a system has been built or modified. Managers must therefore predict how a proposed layout, capacity level, staffing plan, material-handling arrangement, or service configuration will perform under uncertain conditions. Static calculations are useful for estimating capacity and space requirements, but they may not reveal the dynamic effects of variability, congestion, blocking, starvation, rework, breakdowns, or changing demand patterns.

Simulation and analytics provide a bridge between facility design and operational performance. They enable decision-makers to represent the flow of entities through a system, observe the interaction of resources and queues, and compare alternative designs before committing capital. In a production facility, the entities may be components, jobs, pallets, vehicles, or customers. In a hospital, they may be patients. In a warehouse, they may be orders or storage movements.

This chapter examines discrete-event simulation, queueing analysis, scenario analysis, Monte Carlo methods, performance dashboards, and managerial decision-making. The central principle is that analytical tools do not replace managerial judgment. They improve judgment by making assumptions explicit, exposing trade-offs, and providing evidence about the likely consequences of competing alternatives.

Learning Objectives

After studying this chapter, the learner should be able to:

1. Explain the role of simulation and analytics in facility planning.

2. Distinguish discrete-event simulation, queueing models, Monte Carlo analysis, and scenario analysis.

3. Develop a conceptual model of a facility system and identify its principal inputs, resources, queues, and outputs.

4. Interpret utilization, throughput, waiting time, work-in-process, cycle time, and service-level measures.

5. Design experiments that compare facility alternatives under uncertainty.

6. Recognize issues of validation, data quality, randomness, model risk, and statistical confidence.

7. Use dashboards and analytical evidence to support managerial facility decisions.

8. Evaluate the ethical, organizational, and governance implications of simulation-based recommendations.

From Static Layouts to Dynamic Facility Models

Traditional layout analysis often evaluates distances, adjacency, space, and estimated capacity. These dimensions remain important, but facility performance is dynamic. A work centre may have adequate nominal capacity yet become a bottleneck because arrivals are uneven, processing times vary, batches are released simultaneously, or a downstream operation fails. Similarly, a layout with a short average travel distance may produce poor performance if traffic intersections create repeated congestion.

A dynamic facility model represents the movement and transformation of entities over time. It normally includes:

• entities, such as jobs, orders, vehicles, or patients;
• activities, such as processing, inspection, storage, transport, or service;
• resources, such as machines, workers, docks, or automated vehicles;
• queues and buffers;
• routing rules and priorities;
• failures, repairs, setups, and other interruptions;
• arrival and demand patterns; and
• performance measures.

The model must be detailed enough to capture the mechanisms that influence the decision, but not so detailed that it becomes impossible to build, validate, or interpret. This is the principle of purposeful abstraction. A model designed to compare forklift routes may require traffic rules and travel-time distributions, whereas a model designed only to estimate total annual capacity may not require individual vehicle movements.

Core Analytical Concepts

Utilization and capacity

Utilization is the proportion of available resource time used for productive or specified activity. If a machine is busy for 420 minutes during a scheduled 480-minute shift, its observed utilization is 87.5 per cent. Utilization is not equivalent to efficiency. High utilization may indicate good asset use, but when a resource operates near full capacity, small disturbances can create disproportionately large queues and delays. Practical capacity must therefore include allowances for setups, maintenance, absenteeism, quality losses, and variability.

Throughput and flow time

Throughput is the rate at which completed entities leave the system. Flow time, or cycle time, is the elapsed time from entry to completion. A facility may increase throughput while worsening customer flow time if it releases excessive work into the system. This is why throughput should be interpreted together with work-in-process and waiting time.

Work-in-process and Little’s Law

Little’s Law states:

WIP = Throughput × Average Flow Time

When the system is stable and the measures are defined consistently, the relationship provides both a diagnostic tool and a managerial insight. If throughput remains constant but average flow time rises, WIP must also rise. A facility that seeks shorter lead times cannot simply increase releases; it must reduce waiting, variability, processing time, or constraints.

Bottlenecks and constraints

A bottleneck is a resource or condition that limits system output. It may be a machine, labour skill, inspection station, dock, aisle, policy, information system, or approval step. The bottleneck can move when the system is changed. For this reason, improving one station in isolation may have little effect if another constraint becomes dominant.

Service level

Service level may be defined as the proportion of orders completed before a promised time, the probability that demand is met without stockout, or the percentage of customers served within a specified waiting-time target. Averages can conceal poor service. Percentiles, such as the 95th-percentile waiting time, are often more useful for decisions involving customer promises and risk.

Discrete-Event Simulation

Discrete-event simulation, or DES, models a system as a sequence of events occurring at particular points in time. An event changes the state of the system. Examples include job arrival, processing completion, machine failure, vehicle arrival, order release, inspection rejection, or repair completion.

A basic DES progresses as follows:

1. The simulation clock advances to the next scheduled event.

2. The state of the system is updated.

3. Any queues, resource statuses, and entity attributes affected by the event are changed.

4. New events are scheduled.

5. Relevant performance statistics are recorded.

The model may use a first-in-first-out queue, priority rules, shortest-processing-time sequencing, appointment rules, or dispatching policies. It may also include batching, rework, alternative routes, resource sharing, and finite buffers.

Building a useful simulation model

Model development begins with a clearly framed decision question. “Should the plant be expanded?” is too broad. A better question is: “Will adding one inspection station reduce the 95th-percentile order lead time sufficiently to justify its cost under expected demand growth?”

The analyst then defines the system boundary, entities, processes, resources, routes, assumptions, input data, outputs, and experimental alternatives. A process map or value-stream map is often useful for communication. Subject-matter experts should review the conceptual model before coding begins.

Input data may include interarrival times, processing times, setup durations, failure rates, repair times, transport times, demand mix, staffing schedules, and routing probabilities. Data should be examined for seasonality, trends, outliers, autocorrelation, and changes in operating policy. A single average value can be misleading. For example, using an average processing time ignores the operational impact of long jobs if those jobs generate substantial queues.

Verification and validation

Verification asks whether the model has been implemented correctly. It involves checking logic, units, event sequencing, conservation of flow, resource calendars, and extreme cases. Animation can help identify routing or congestion errors, but visual appearance is not proof of validity.

Validation asks whether the model is a credible representation of the real system for the intended purpose. Historical performance, expert review, face validity, and comparison with known operating conditions may be used. A model need not reproduce every historical observation exactly; it must reproduce the relationships and performance measures relevant to the decision.

Warm-up, replications, and confidence

A simulation that starts empty may initially appear to perform better than a system operating continuously. The warm-up period allows the model to reach representative conditions and is excluded from performance estimates when appropriate. Independent replications with different random-number streams are then used to estimate mean performance and variation. Results should be presented with confidence intervals rather than as falsely precise single values.

For example, if ten replications estimate average flow time as 6.8 hours with a 95 per cent confidence interval of 6.4 to 7.2 hours, the decision-maker should recognize both the expected value and the uncertainty. If two alternatives have overlapping and wide intervals, further analysis or better data may be necessary.

Queueing Analysis as a Structural Lens

Queueing theory provides mathematical approximations for systems in which entities arrive, wait, receive service, and depart. A queueing model is often represented using notation such as M/M/1: Markovian or Poisson arrivals, exponentially distributed service times, and one server. More elaborate models may include multiple servers, finite capacity, priority classes, or general service distributions.

For a simple stable single-server queue, utilization is represented by ρ = λ/μ, where λ is the arrival rate and μ is the service rate. As ρ approaches one, expected waiting time increases sharply. This explains why a resource operating at 98 per cent utilization may produce much longer delays than one operating at 85 per cent, even though the difference in utilization appears modest.

Queueing models are valuable for rapid screening, sensitivity analysis, and insight into the relationship between demand, capacity, and delay. Their assumptions may be too restrictive for complex facilities with finite buffers, routing interactions, breakdowns, and multiple product types. DES is generally more flexible, while queueing models are often faster and more transparent. The two approaches should be viewed as complementary rather than competing.

Scenario Analysis and Monte Carlo Methods

Scenario analysis compares defined alternatives under specified assumptions. Scenarios may represent demand levels, staffing patterns, layout options, automation strategies, product mixes, or disruption conditions. A useful scenario is not merely an optimistic, most-likely, or pessimistic label; it specifies the operational assumptions that generate the outcome.

Monte Carlo analysis repeatedly samples uncertain inputs from probability distributions and calculates the resulting output. It is useful when the decision-maker wants a distribution of possible outcomes rather than one forecast. For example, annual demand, processing time, equipment availability, and supplier lead time may all be uncertain. The analysis can estimate the probability that throughput falls below a target or that a proposed layout exceeds its capacity in a peak month.

Care is required when assigning distributions. A distribution should reflect data and operational knowledge. Correlations must also be considered. Demand and staffing may rise together during peak periods, while machine failures may increase with utilization. Treating correlated variables as independent can understate risk.

Sensitivity analysis identifies which assumptions have the greatest effect on results. Tornado charts, elasticity measures, and one-factor or multi-factor experiments can reveal whether a recommendation depends primarily on demand, processing time, labour availability, or equipment reliability. This directs managerial attention toward the uncertainties worth reducing.

Designing Facility Experiments

A sound experiment compares alternatives on a common basis. The alternatives should use the same demand realizations, where appropriate, so that differences are attributable to the facility design rather than unrelated random variation. This technique, known as common random numbers, can improve the precision of comparisons.

Typical performance measures include:

• average and percentile flow time;
• throughput per shift or day;
• WIP by process and in total;
• resource utilization;
• queue length and waiting time;
• on-time completion or service level;
• travel distance and material-handling effort;
• labour hours, overtime, and energy use;
• blocking and starvation frequency; and
• cost, revenue, or return on investment.

No alternative should be judged on a single metric. A smaller footprint may increase travel distance. Higher utilization may reduce spare capacity and resilience. Automation may increase throughput but require specialized maintenance. Multi-criteria evaluation makes these trade-offs explicit.

A practical experiment often proceeds through screening, refinement, and stress testing. Screening removes clearly inferior options. Refinement tests layout dimensions, staffing levels, buffer sizes, dispatching rules, or equipment quantities. Stress testing examines peak demand, absenteeism, breakdowns, supply interruptions, and product-mix changes.

Dashboards and Decision Communication

A dashboard translates analytical results into a form that managers can monitor and discuss. An effective facility dashboard connects strategic objectives to operational measures. It should distinguish leading indicators, such as queue growth and preventive-maintenance compliance, from lagging indicators, such as missed deliveries and actual throughput.

A dashboard should show definitions, time periods, targets, data sources, and appropriate levels of aggregation. Colour coding should not conceal uncertainty or differences in product mix. Drill-down capability can allow a manager to move from total order lead time to a process, shift, resource, or product family.

Simulation outputs should be labelled as estimates, not actual observations. A dashboard may display a predicted 90th-percentile lead time, a confidence interval, and the assumptions under which the prediction was generated. Governance is strengthened when model versions, data refresh dates, and ownership are visible.

Illustrative Case: Redesigning a Medical-Device Assembly Facility

Background

The following is an illustrative case developed for educational purposes. A medical-device manufacturer operates a high-mix assembly facility. Orders pass through kitting, manual assembly, inspection, packaging, and dispatch. Management is considering two alternatives: adding an inspection station or redesigning the layout so that assembly and inspection are placed closer together.

Situation and problem

Average monthly throughput meets demand, but customer service is deteriorating. The average order lead time is acceptable, while the 90th percentile is substantially higher. Operators report that inspection queues grow after large batches are released. Material handlers also make repeated trips between assembly and inspection. Management wants to determine whether the problem is insufficient inspection capacity, poor release control, or excessive movement.

Analysis

The project team defines entities as customer orders and models the five process stages. Processing times are represented by empirical distributions by product family. The model includes batch releases, inspection rework, finite staging space, two shifts, planned breaks, equipment downtime, and the existing dispatching rule. Historical output is used for validation, and supervisors review the routing logic.

The baseline model shows that inspection utilization is high during release peaks, while assembly is intermittently starved. The average utilization figures alone do not reveal the problem. Queue accumulation at inspection creates congestion, which delays the return of inspected work and causes assembly staging areas to fill. The system also experiences blocking when the packaging area cannot accept completed orders.

Alternatives

Four alternatives are tested:

1. Add one inspection station.

2. Relocate inspection beside assembly without adding capacity.

3. Add a station and introduce smaller, controlled release batches.

4. Relocate inspection, introduce controlled release, and establish a cross-trained relief operator.

Each alternative is tested under normal demand, a peak-demand scenario, and a scenario with elevated equipment downtime. Measures include throughput, 90th-percentile lead time, WIP, inspection utilization, travel distance, overtime, and service level.

Findings and decision approach

The additional station increases nominal capacity, but its benefit is limited when large batches continue to arrive. Relocation reduces travel and improves visibility but does not fully resolve inspection queues. The combined alternative performs best because it addresses both capacity and flow variability. Its advantage is greatest in the peak-demand scenario. The cross-trained operator also improves resilience during absenteeism and downtime.

The recommendation is not based only on the best average lead time. Management considers implementation cost, training requirements, validation of the new inspection arrangement, and the risk that demand will shift toward a product family with longer inspection times. A staged implementation is selected: controlled release is introduced first, followed by layout relocation and cross-training. The model is retained as a monitoring tool and updated after operational data become available.

Learning points

The case illustrates four principles. First, the visible queue may be a symptom of release variability rather than a simple shortage of resources. Second, a layout change and a control-policy change can be mutually reinforcing. Third, average performance may hide service-risk problems that appear in the upper tail of the distribution. Fourth, simulation is most useful when linked to an implementation plan and subsequent measurement.

Managerial Applications and Decision Governance

Facility analytics can support location and layout selection, dock sizing, warehouse slotting, staffing, buffer design, equipment acquisition, automation justification, preventive-maintenance planning, and capacity expansion. At the strategic level, models can compare network configurations or postponement policies. At the operational level, they can evaluate dispatching rules, shift patterns, and replenishment policies.

Managers should distinguish a model recommendation from a decision. A model evaluates specified assumptions; management decides whether those assumptions are credible, whether the benefits justify the investment, and whether the organization can implement the change. Financial analysis should include capital cost, labour, maintenance, energy, training, disruption during implementation, and the value of improved service.

Governance requires clear ownership of data, assumptions, model changes, and approval authority. Significant models should have documented purpose, scope, version, validation evidence, limitations, and review dates. When analytical tools influence staffing or employee performance evaluation, the organization must ensure that measures are fair, transparent, and not based on incomplete or biased data.

Challenges, Risks and Common Errors

The most common error is false precision. A simulation may produce detailed charts even when the input data are weak. Another error is optimizing the wrong objective, such as maximizing machine utilization while ignoring customer lead time. Analysts may also omit informal work, quality interruptions, changeovers, material shortages, or human recovery actions because these are difficult to model.

Other risks include:

• using averages where variability is operationally important;
• failing to model correlations and peak conditions;
• validating only the visual animation rather than the performance logic;
• running too few replications;
• excluding warm-up effects incorrectly;
• comparing alternatives with inconsistent assumptions;
• treating a forecast as a guarantee;
• ignoring implementation and behavioural responses; and
• allowing dashboards to encourage local optimization.

A model can be technically correct and still be managerially misleading if the decision question, boundary, or performance measures are inappropriate. Periodic review is necessary because facilities evolve. New products, labour practices, demand patterns, and automation technologies can invalidate earlier assumptions.

Technology, Digital Twins and Emerging Practice

Modern simulation platforms increasingly connect with enterprise resource planning, manufacturing execution, warehouse management, Internet of Things, and building-management systems. Real-time data can update equipment status, order queues, travel times, and demand signals. A digital twin extends this idea by maintaining a data-linked representation of a physical facility for monitoring, prediction, and experimentation.

Artificial intelligence and machine learning can assist with demand forecasting, anomaly detection, parameter estimation, and optimization. However, predictive accuracy does not remove the need for causal understanding. A black-box recommendation may be difficult to validate, explain, or govern when major capital decisions are involved. Hybrid approaches, combining process-based simulation with data-driven models, are often more credible for facility decisions.

Emerging applications include autonomous vehicle traffic simulation, energy-aware facility design, human-robot coordination, disruption modelling, and real-time rescheduling. Cybersecurity, data ownership, interoperability, and model drift become increasingly important as analytical systems are connected to operational technology.

Best Practices for Evidence-Based Facility Decisions

A disciplined practice includes the following principles:

1. Begin with a decision and define the measures that matter.

2. Build the simplest model capable of answering the question.

3. Involve operators, engineers, planners, finance specialists, and managers.

4. Use representative data and document assumptions.

5. Validate both model logic and output behaviour.

6. Represent variability, peak conditions, and relevant dependencies.

7. Compare alternatives using common assumptions and multiple replications.

8. Report ranges, percentiles, and confidence intervals rather than unsupported point estimates.

9. Combine operational, financial, human, environmental, and resilience criteria.

10. Pilot important changes and update the model with observed results.

Chapter Summary

Simulation and analytics enable facility planners to evaluate dynamic performance before implementing physical or organizational changes. Discrete-event simulation represents entities, resources, queues, routing, failures, and events over time. Queueing theory provides rapid insight into the relationship between arrival rates, service capacity, utilization, and waiting. Scenario analysis compares defined alternatives, while Monte Carlo methods quantify the effects of uncertain inputs.

Important performance measures include throughput, utilization, flow time, WIP, queue length, service level, travel, cost, and resilience. Little’s Law links WIP, throughput, and flow time, while bottleneck analysis explains system-wide constraints. Reliable analysis requires appropriate data, purposeful abstraction, verification, validation, warm-up treatment, independent replications, confidence intervals, and sensitivity analysis.

Dashboards and digital technologies improve communication and monitoring, but analytical outputs remain dependent on assumptions and governance. Effective facility decision-making therefore combines quantitative evidence with operational expertise, financial reasoning, ethical responsibility, and practical implementation planning.

Key Terms

Discrete-event simulation: A model in which system state changes at distinct events over simulated time.

Entity: An item that flows through a system, such as a job, order, vehicle, or patient.

Resource: A person, machine, space, or equipment unit required to perform an activity.

Queue: A line or waiting area formed when demand for a resource exceeds immediate availability.

Utilization: The proportion of available resource time used for specified activity.

Throughput: The rate at which completed entities leave a system.

Flow time: Total elapsed time from an entity’s entry into a system until completion.

Work-in-process: Entities that have entered a process but have not yet been completed.

Bottleneck: A constraint that limits the output or flow of the system.

Service level: The proportion of demand meeting a defined performance requirement.

Little’s Law: The relationship WIP = throughput multiplied by average flow time.

Queueing theory: Mathematical analysis of arrival, waiting, service, and departure processes.

Monte Carlo analysis: Repeated random sampling of uncertain inputs to estimate output distributions.

Scenario analysis: Structured comparison of system performance under different assumptions or alternatives.

Verification: Checking whether a model has been implemented according to its design.

Validation: Assessing whether a model is credible for its intended purpose.

Warm-up period: Initial simulation time excluded when the model begins in an unrepresentative state.

Replication: An independent simulation run using a different random-number stream.

Confidence interval: A statistical range expressing uncertainty around an estimated performance measure.

Digital twin: A data-connected representation of a physical system used for monitoring and analysis.

Review Questions

Short Answer Questions

1. Why can a facility with high average utilization experience severe congestion?

2. State Little’s Law and explain the meaning of its three variables.

3. What is the difference between verification and validation?

4. Why are percentiles useful in service-level analysis?

5. What is the purpose of a warm-up period in simulation?

Long Answer Questions

1. Explain the process of developing and validating a discrete-event simulation model for a manufacturing facility.

2. Compare queueing models, discrete-event simulation, scenario analysis, and Monte Carlo analysis.

3. Discuss how bottlenecks, variability, and release policies interact to influence WIP and flow time.

4. Explain how dashboards can communicate simulation results without creating false precision.

5. Discuss the managerial, ethical, and governance issues involved in using analytics for facility decisions.

Analytical / Application Questions

1. A work centre receives 48 jobs per shift and completes them at the same average rate. Its average flow time rises from 2 hours to 3.5 hours. Use Little’s Law to explain the likely change in WIP and identify operational causes that should be investigated.

2. A proposed automated system increases throughput by 12 per cent but raises capital cost, maintenance skill requirements, and dependence on a single technical specialist. Design a multi-criteria evaluation using simulation outputs and managerial measures.

3. A warehouse meets its average daily demand but misses delivery promises during two weekly demand peaks. Propose a simulation experiment, including inputs, scenarios, performance measures, validation steps, and the decision rule for selecting an intervention.

22. Multiple Choice Questions with Answers

Question 1

Which measure is most directly concerned with the proportion of orders completed before a promised time?

A. Utilization

B. Service level

C. Travel distance

D. Resource availability

Correct Answer: B. Service level

Question 2

A simulation begins empty, although the real facility operates continuously. Which technique helps prevent initial conditions from distorting results?

A. Increasing utilization

B. Using only one replication

C. Applying a warm-up period

D. Removing all queues

Correct Answer: C. Applying a warm-up period

Question 3

According to Little’s Law, if throughput remains constant and average flow time increases, what normally happens to WIP?

A. It decreases

B. It remains zero

C. It increases

D. It becomes unrelated to throughput

Correct Answer: C. It increases

Question 4

What is the primary purpose of model validation?

A. To make the animation attractive

B. To determine whether the model is credible for its intended use

C. To eliminate all variability

D. To guarantee the recommended outcome

Correct Answer: B. To determine whether the model is credible for its intended use

Question 5

Why can a resource operating at 98 per cent utilization create disproportionately large delays?

A. Waiting generally rises sharply as utilization approaches full capacity

B. High utilization eliminates all queues

C. Utilization measures only product quality

D. High utilization always reduces processing time

Correct Answer: A. Waiting generally rises sharply as utilization approaches full capacity

Question 6

Which method is most suitable for estimating the probability that a facility misses its throughput target under uncertain demand and processing times?

A. Monte Carlo analysis

B. A fixed average calculation only

C. Adjacency analysis

D. A break-even chart without uncertainty

Correct Answer: A. Monte Carlo analysis

Question 7

In the illustrative medical-device case, why did controlled release improve the benefit of additional inspection capacity?

A. It eliminated the need for inspection

B. It reduced the variability and batching of arrivals to inspection

C. It increased travel distance

D. It made all processing times identical

Correct Answer: B. It reduced the variability and batching of arrivals to inspection

Question 8

Which practice improves the fairness of comparing two stochastic simulation alternatives?

A. Using different demand assumptions for each alternative

B. Comparing only the best replication

C. Using common random numbers where appropriate

D. Ignoring confidence intervals

Correct Answer: C. Using common random numbers where appropriate

Question 9

A model shows excellent average lead time but a very poor 95th-percentile lead time. What does this most likely indicate?

A. All orders have identical performance

B. A subset of orders experiences significant delay

C. Throughput must be zero

D. The model contains no variability

Correct Answer: B. A subset of orders experiences significant delay

Question 10

Which statement best describes a digital twin in facility planning?

A. A static floor plan with no operational data

B. A data-connected representation of a physical system used for monitoring and analysis

C. A financial forecast independent of operations

D. A replacement for all managerial judgment

Correct Answer: B. A data-connected representation of a physical system used for monitoring and analysis