Simulating a production line before you build it: buffers, variability and bad inputs

Why coupled lines underperform the arithmetic

Connect several stations directly and the line stops whenever any of them stops, so combined availability falls well below the weakest station's individual figure. Add variability in cycle time and the effect compounds, because a station that finishes early cannot bank the time but a station that runs long delays everything behind it. This is why a line built from individually capable machines produces less than expected and nobody can point to the machine at fault. Simulation makes the interaction visible and lets you test the two remedies — buffering between operations and reducing stoppage frequency — before committing capital to either.

Buffer sizing is the question simulation answers best

Buffers decouple stations so a short stoppage at one does not immediately starve or block its neighbours. Too small and they achieve nothing; too large and they consume floor space, tie up work in progress and lengthen the time between making a defect and discovering it. The right size depends on the distribution of stoppage durations, not the average, because occasional long stoppages behave completely differently from frequent short ones. That distribution is precisely what a spreadsheet cannot handle and a simulation can, provided the durations fed into it came from measurement rather than from an engineer's recollection.

Inputs decide the answer, and the inputs are usually guesses

A study needs cycle times by station and variant, changeover durations, stoppage frequency and duration distributions, scrap and rework rates, and staffing rules. Most plants have none of this to the accuracy required, so numbers get assumed and the output inherits their errors while looking authoritative. Spend the effort on measurement first: instrument the existing line, or time the operations, and be explicit in the report about which inputs were measured and which were assumed. Where an input is uncertain, run the model across its plausible range and see whether the recommendation changes, because sometimes it does not and the uncertainty stops mattering.

Use it to compare options, not to predict a number

The reliable output of a simulation is a ranking: this layout beats that one, this buffer position matters more than the other, the constraint moves here when volume rises. The unreliable output is an absolute figure quoted to a business case as if it were a measurement. Frame studies as comparisons under identical assumptions, since errors common to both options largely cancel, and state the conclusion in terms of relative behaviour and sensitivity. Presenting a single output figure invites a debate about the model's credibility and obscures the finding that actually deserved attention.

Where a model earns its keep after commissioning

Most models are built for one capital decision and then discarded, which wastes the effort of assembling the data. A maintained model answers recurring operational questions: how a new variant affects flow, whether an extra shift on one station relieves the constraint, what happens when a machine is taken out for a rebuild, how a demand increase redistributes the bottleneck. Keeping it alive requires an owner, updated data and a licence, so decide at the outset whether that is realistic. If not, extract and record the insight rather than pretending the model will be reused.

Frequently asked questions

Is simulation worth it for a small line?

It depends on coupling and cost of error, not on size. A short line with tightly linked stations, meaningful variability and a large capital commitment benefits from a model. A few independent machines with buffers between them can usually be understood well enough by hand. The other consideration is whether the exercise of gathering the input data would itself be valuable, because plants often learn more from measuring their real stoppage patterns than from the study those measurements feed.

What input data do we need before commissioning a study?

Station cycle times for each variant, changeover times, the frequency and duration of stoppages rather than a single availability percentage, scrap and rework behaviour, and the rules people actually follow when something goes wrong. The distribution of stoppage duration matters more than its average. If that data does not exist, gathering it is the first phase of the project, and any study run without it should be presented as a comparison of options rather than as a forecast.

Can simulation results be trusted in a capital submission?

As comparative evidence, yes, provided the assumptions are stated and the sensitivity to uncertain inputs is shown. As a precise output prediction, treat with caution, because the model reflects the behaviours somebody chose to represent and real lines fail in ways nobody modelled. The strongest submissions present the ranking of options, the range of outcomes across plausible inputs, and an explicit list of what the model does not cover, such as quality escapes or material supply interruptions.