Value Stream Simulation: Make Variability, Buffers and Bottlenecks Visible
A value stream map works with averages: one cycle time per process, one inventory count per location. In reality, times vary, machines break down, and batches wait in front of the bottleneck. A simulation runs piece by piece through the map and shows what the averages hide.
What the static map misses
The value stream timeline adds up processing times and the days of supply in each inventory. It tells you how long a piece takes today on average, but not how much that varies, whether a buffer runs empty, or why more and more parts pile up in front of the bottleneck. Three effects are missing: variability of cycle times, the coupling of processes through limited buffers, and the randomness of downtime and scrap.
The Kingman formula describes the first effect: waiting time in front of a station grows disproportionately with utilization and with the square of variability. At 90% utilization the utilization factor is 0.9 ÷ 0.1 = 9. With exponentially distributed arrivals and processing (c = 1), a piece waits on average nine times the processing time; with moderate variability (c = 0.5) it still waits 2.25 times.
How the simulation reads the map
ValueStreamMapping.app turns the main path into a chain of stations and buffers. Every process with a cycle time becomes a station with its number of machines, uptime, scrap and batch size. Every inventory symbol in between becomes a buffer: FIFO lanes are capped at their maximum, supermarkets at their target inventory, warehouses are unlimited, and two processes with no symbol between them run as one-piece flow.
Cycle times vary lognormally around their mean, breakdowns follow the uptime with a mean repair time, and scrap hits each piece with its probability. Production control releases orders at takt time, and the customer pulls from finished goods at takt time. If nothing is there, a backlog builds up. Several replications with different random numbers show the range of results (Monte Carlo).
- Output per day against demand: does the line meet takt time?
- Service level: share of customer requests filled immediately, and the largest backlog.
- Lead time as a distribution: mean and P95 against the static timeline.
- Inventory over time per buffer: is a warehouse growing, a supermarket running empty, a FIFO lane full?
- Time shares per process: processing, changeover, down, blocked (next buffer full), starved (previous buffer empty).
Reading the results
A process that spends close to 100% of its time processing is the bottleneck; the buffer in front of it grows, the buffers after it run empty. "Blocked" at a fast upstream process means the FIFO lane or supermarket is too small or the downstream process is too slow. "Starved" at a downstream process points to breakdowns or batches that are too small upstream.
Little's Law connects the three metrics: work in process = throughput × lead time. If inventory drops at the same throughput, lead time drops by the same ratio. The simulation shows how far an inventory can go down before the service level collapses.
What-if analysis
Instead of building variants by hand, the analysis tests for each process what you gain from a 10% shorter cycle time, an extra machine, full uptime, zero scrap or half the batch size, and for each limited buffer what double capacity brings. All variants run with the same random numbers as the baseline, so small differences do not get lost in the noise. The table is sorted by the gain in service level and output.
The best variant is your next improvement. Often it is not the bottleneck itself but the breakdowns at the process before it, or a FIFO lane that is too small and starves the bottleneck.
Practice in the editor
Open the template "Steering bracket (current state)", switch to the "Simulation" tab and run the simulation. Find the bottleneck, start the playback on the map and run the what-if analysis. Which single change raises the service level the most?
Open the template in the editor Start blankFrequently asked questions
How long should the simulation run?
At least three times the static lead time plus two days of warm-up, so that pieces started after the warm-up also reach the end. ValueStreamMapping.app suggests this run length.
What variability is realistic?
For manual work a coefficient of variation of 20 to 30%, for machines 5 to 15%. If you have your own readings from the stopwatch data collection, divide the standard deviation by the mean.
Why does an inventory grow endlessly in the simulation?
Because the process after it is slower than takt time (bottleneck above takt) and still receives material under push. In reality, production control steps in at some point; the map shows that the bottleneck and order release do not match.
Does this replace a factory simulation such as Plant Simulation?
No. It models the main path of the value stream with the data from the map, without layout, routings per product or shift calendar. In return, it is ready in seconds and needs no extra modeling.