The Zipper Merge, Explained

After reading this you will know why the late (zipper) merge usually moves more cars than early merging, how the Intelligent Driver Model turns individual braking into road-wide shockwaves, and how to read throughput and delay off the simulator without fooling yourself.

What the simulator models

A two-lane road drops to one lane. Every car must end up in the surviving lane before the taper. The moral argument is old: some drivers merge early and queue in a single file, others ride the closing lane to the very end and slot in one at a time. This tool does not argue. It runs the cars forward under a car-following rule and measures the outcome.

Here is the hook. Suppose 2000 cars per hour arrive at a bottleneck that can pass only 1500 per hour. A queue must form either way, because more metal wants through than the road allows. The merge style cannot raise that ceiling of 1500. What it can do is decide how the queue stores itself and whether panic braking at the merge point drops the ceiling below 1500. That second effect is where the whole debate lives.

Each vehicle follows the Intelligent Driver Model (IDM). Given the gap to the car ahead and the speed difference, IDM outputs an acceleration. Feed that acceleration back into position and velocity every time step, and stop-and-go waves, backward-travelling shockwaves and capacity drop all appear on their own. You do not script the jam. It emerges.

When the answer matters and when it does not

Merge style only changes the outcome when the road is congested, meaning arrivals exceed the bottleneck capacity for a sustained stretch. In free-flowing traffic every car reaches the merge at speed, slots in, and merge style barely registers.

Under congestion two things change. First, the late merge uses both lanes as queue storage, so a queue of the same number of cars is roughly half as long. That matters when the queue would otherwise spill back into an upstream junction. Second, a messy merge with drivers jockeying and slamming brakes lowers the effective capacity below its clean value. A courtesy zipper (strict one-for-one alternation) keeps merge speed high and prevents the lane-blocking standoffs.

The clean-flow capacity of a lane drop is set by the physics of the bottleneck, not by manners. Merging style mostly redistributes delay and queue length. The one lever it has on throughput is avoiding the capacity drop that panic braking causes.

The Intelligent Driver Model in one equation

IDM gives the acceleration of a car from its own speed v, the gap s to the car ahead, and the approach rate \Delta v (its speed minus the leader's speed).

\dot v = a \left[ 1 - \left( \frac{v}{v_0} \right)^{4} - \left( \frac{s^{*}(v, \Delta v)}{s} \right)^{2} \right]

Here a is the maximum comfortable acceleration, v_0 is the driver's desired free speed, and s is the bumper-to-bumper gap. The first bracketed term drives the car toward v_0 when the road is clear. The second term brakes when the car is closer than its desired gap.

The desired gap s^{*} is itself a function:

s^{*}(v, \Delta v) = s_0 + v\,T + \frac{v\,\Delta v}{2\sqrt{a\,b}}

Now s_0 is the jam gap (the space you keep when stopped), T is the desired time headway, and b is the comfortable braking deceleration. When you approach fast, \Delta v is large and positive, so s^{*} grows and the car demands more braking. When speeds match, that last term vanishes and the car settles to a steady following gap of s_0 + vT.

Typical values used here: v_0 = 30 m/s, a = 1.5 m/s², b = 2.0 m/s², T = 1.2 s, s_0 = 2 m, and a car length of 5 m.

Why capacity has a ceiling you can compute

At steady speed the equilibrium gap is s = s_0 + vT. The distance each car occupies is that gap plus one car length L. The flow (cars per second) is speed divided by occupied distance.

Q(v) = \frac{v}{L + s_0 + vT}

Plug in the numbers. At v = 30 m/s, occupied distance is 5 + 2 + 30·1.2 = 43 m, so Q = 30/43 = 0.698 cars/s, about 41.9 cars per minute in that lane. At a crawling v = 5 m/s, occupied distance is 5 + 2 + 6 = 13 m, giving Q = 5/13 = 0.385 cars/s, about 23.1 per minute. Flow is highest at some intermediate speed, not at a standstill and not at full speed.

Flow rises with speed and flattens near free speed. Because the numerator and denominator both grow with v, throughput approaches a ceiling rather than climbing forever.

The correct reading: throughput keeps climbing with speed toward a ceiling of 1/T = 1/1.2 = 0.833 cars/s = 50 cars/min in the limit. But real congestion sits well below free speed. The lesson that matters is the capacity drop: once flow breaks down into stop-and-go, the discharge rate at the bottleneck falls, and it stays low even after arrivals ease. Keeping merge speed high is how you avoid falling off the top of this curve.

A worked example reproducing the demo

Early merge versus courtesy zipper at 1800 cars/hour

Run the demo defaults: arrivals at 1800 cars/hour (30 cars/min) split across two lanes, a lane drop, and the courtesy zipper toggle off. Compare that against turning courtesy on. The single surviving lane has a clean discharge capacity of about 25 cars/min at the merge crawl speed of roughly 6 m/s.

Arrivals are 30 cars/min. Bottleneck clean capacity is about 25 cars/min. Demand exceeds capacity by 5 cars/min, so a queue must grow.
Early merge, courtesy off. Everyone funnels into one lane far upstream. The queue is long and single-file. Because merges happen calmly upstream, discharge holds near 25 cars/min. Delay per car settles around 48 s once the queue reaches steady length.
Late merge, courtesy off. Drivers hold both lanes to the taper, but with no alternation rule they jockey and brake hard. Discharge drops to about 21 cars/min (a capacity drop of roughly 16 percent). Queue is shorter but slower to clear, and delay climbs toward 62 s.
Late merge, courtesy on. Strict one-for-one alternation keeps merge speed up. Discharge recovers to about 25 cars/min, queue length halves versus early merge, and delay settles near 47 s.

The verdict from these numbers: the zipper wins on queue length in every case and ties or beats early merging on delay only when alternation is enforced. A disorderly late merge is the worst of the three because it triggers the capacity drop.

Discharge and delay side by side. The courtesy zipper matches early-merge throughput while cutting queue length; the disorderly late merge loses on both.

Reading and interpreting the results

Two readouts carry the argument: throughput in cars per minute and average delay per car. Colour shows speed, so a red band marks where cars crawl.

Throughput: Cars crossing the bottleneck per minute. Under congestion this equals the discharge rate. Compare it to the clean capacity to see if a capacity drop occurred.
Average delay: Extra travel time per car versus free-flow. It rises as the queue grows and is the number you should optimise, not queue length alone.
Queue length: How far back the slow band reaches. Late merging roughly halves this because both lanes store cars.
Shockwave: The boundary between fast and slow traffic. It travels backward at roughly 15 to 20 km/h regardless of how fast cars move.

Watch for the moment flow breaks down. Before breakdown the surviving lane runs near 25 cars/min. After a hard merge, it can sit at 21 and refuse to recover even when the incoming rate falls. That hysteresis is the capacity drop, and it is the single most important thing the simulator teaches.

Common mistakes

Do not conclude the zipper is faster in all traffic. Below capacity, both strategies give near-zero delay and the comparison is noise. The zipper's real gain is halved queue length plus protection against capacity drop, and that only appears under sustained congestion.

A second mistake is comparing queue length instead of delay. A short dense queue and a long thin queue can hold the same number of cars. What you care about is how fast the bottleneck discharges them, which is throughput and delay.

A third mistake is treating one run as truth. IDM with a slightly random arrival stream will vary run to run. Average several runs at the same settings before you trust a 2 cars/min difference.

A fourth: the model has no lane-change aggression parameter beyond what the courtesy toggle encodes, and no trucks, no rubbernecking, no phones. It shows the merge mechanism cleanly, which is exactly why it cannot predict a specific real interchange to the car.

Related simulations

If the emergent-jam physics interests you, the Traffic Jam Simulator shows phantom jams forming from nothing on a ring road with a simpler cellular rule. For merging and routing at a larger scale, try the City Traffic Grid, and for a counterintuitive network result see Braess's Paradox, where adding a road makes everyone slower. Queue theory in a non-traffic setting appears in the Queue Simulator, and the same emergent-from-simple-rules flavour drives the Crowd Evacuation Simulator.

Frequently asked questions

Is the zipper merge actually faster?

Under congestion, a courtesy zipper matches early-merge throughput (about 25 cars/min in the demo) while roughly halving queue length. A disorderly late merge is worse than both because jockeying triggers a capacity drop, cutting discharge to about 21 cars/min. So the honest answer is: yes when alternation is enforced, no when the merge is chaotic.

Why do traffic authorities recommend late merging?

Because it uses both lanes as queue storage. A queue of a given number of cars is about half as long, so it is less likely to block an upstream junction. Germany and several US states now sign the zipper merge explicitly in congestion.

Can merge style raise the road's capacity?

No. The bottleneck's clean capacity (around 1/T cars per second per lane) is a physical ceiling. Merge style can only avoid lowering it through panic braking. When demand is below capacity, style makes almost no measurable difference.

What is a capacity drop?

Once free flow breaks into stop-and-go, the rate at which the bottleneck discharges cars falls and stays low even after arrivals ease. In the demo it appears as discharge dropping from 25 to about 21 cars/min after a hard merge. It is the reason smooth alternation matters.

Why do jams travel backward?

Each car brakes slightly after the one ahead, so the boundary between fast and slow traffic moves upstream while individual cars move forward. In IDM this backward shockwave travels at roughly 15 to 20 km/h no matter how fast the cars themselves go.

Try it yourself in the tool