A slime mold that designs railways
Physarum polycephalum is a single-celled slime mold with no brain. It forages by spreading a web of tubes between food sources. Tubes that carry a lot of flow thicken; tubes that carry little shrink and disappear. After a few hours, what is left is a network that is short, efficient, and still survives a broken link.
In 2010, Atsushi Tero, Toshiyuki Nakagaki and colleagues put oat flakes on a wet plate in the positions of the cities around Tokyo and let the mould grow (Tero et al., Science 327, 2010). They used light to discourage it from crossing mountains and the sea, the way terrain discourages railway builders. The network it grew looked a lot like the real Greater Tokyo rail network, and they measured it:
| Metric | Tokyo rail | Physarum |
|---|---|---|
| Total length ÷ minimum spanning tree (cost) | ≈ 1.8 | 1.75 ± 0.30 |
| Mean path ÷ minimum spanning tree path (efficiency) | 0.85 | 0.85 |
| Share of links whose failure cuts off part of the network | 4 % | 14 ± 4 % |
They then wrote down the rule the mould follows and turned it into a mathematical model. That model is what I used here. Andrew Adamatzky later repeated the plate experiment for motorways in 14 regions of the world (Bioevaluation of World Transport Networks, 2012); Colombia was not one of them.
Bogotá is a good candidate. It is one of the largest cities in the world without a metro, its first line is under construction, and a second and third line plus a regional train are planned. So: what would the mould build?
The model
The Tero model treats the network as pipes. Every pipe has a conductivity D. At each step:
- Pick two food sources at random. One pushes a fixed amount of flow in, the other takes it out.
- Solve for the pressures, so flow through each pipe is Q = D / L × Δp.
- Grow every pipe by how much it was used and let it decay: dD/dt = |Q|^γ / (1 + |Q|^γ) − D, with γ = 1.8 as in the paper.
Repeat thousands of times and only the useful pipes survive.
Try it on a small grid with a lake in the middle. Five food sources sit north of the lake, so the southern shore is always the long way round. With γ = 1.8 the shore dies and one route wins. Drop γ below 1 and the mould keeps both.
1Pick a pair
Two food sources at random: one pumps, one drains.
+I0/−I0I0 = 2.02Solve the flow
Pressures follow from Kirchhoff’s law; each tube carries
Q =DL· Δp3Grow or decay
Busy tubes thicken, idle tubes fade:
dDdt=|Q|γ1 + |Q|γ− Dγ = 1.8
- Source (+I₀)
- Sink (−I₀)
- Food
- Lake
Tubes are drawn thicker the higher their conductivity D, relative to the strongest tube. With γ > 1 the mould commits to one route and lets the long way round the lake die; with γ < 1 it keeps several routes alive. Change the sliders while it runs.
For Bogotá:
- The dish is a triangular mesh, 0.6 km between points, inside the boundaries of the 19 urban localidades plus Soacha, Mosquera, Funza and Chía (about 960 km², 3,064 points).
- The light is land use. Crossing urban land costs 1; crossing open land, wetlands or the hills costs 2.5. Tero used light for the same job.
- The food is 23 population centres (each localidad and municipality, weighted by population) and 7 job centres: Centro, Centro Internacional, Calle 93, Salitre, Zona Franca, El Dorado and Corabastos.
- Demand: Tero's mould has no idea how many people live in each city. I added one change: pairs are still chosen at random, but the flow they push is scaled by the square root of the product of their weights. Bigger places push harder.
I fixed every parameter before overlaying the planned network, and ran three random seeds. The code and data sources are on GitHub.
Press play to watch the network form. Early on, the mould covers most of the city; by the end, a few thick corridors carry almost everything.
What it built
In the run shown above (seed 1), these corridors survive:
- A northern line: Suba to Chía, along the corridor where Regiotram del Norte is proposed.
- A western diagonal: Suba and Engativá down through Fontibón, Kennedy and Corabastos to Bosa and Soacha. It ties the most populated localidades together without passing through the centre.
- A central spine: Salitre through Teusaquillo to Centro Internacional and Centro.
- A southern loop: Ciudad Bolívar, Tunjuelito and Rafael Uribe to the centre, with a spur to Usme.
- A western link: Mosquera to El Dorado and Zona Franca, close to the Regiotram de Occidente route.
- North-east links: Suba to Usaquén and Calle 93.
The other two seeds grow a similar network: 70 % to 90 % of each seed's strongest 88 km lies within 750 m of another seed's.
Measured the way Tero measured Tokyo, the Bogotá network lands in the same range:
| Metric | Tokyo Physarum | Bogotá Physarum (3 seeds) |
|---|---|---|
| Cost (length ÷ minimum spanning tree) | 1.75 ± 0.30 | 1.54 – 1.66 |
| Efficiency (mean path ÷ minimum spanning tree path) | 0.85 | 0.84 – 0.87 |
| Share of links whose failure cuts off part of the network | 14 ± 4 % | 24 – 31 % |
It is less fault tolerant than Tokyo's mould. In every seed, Chía, Soacha, Usme, Mosquera and Funza, plus a few small central localidades, hang on a single branch: one break cuts them off.
Compared with the plan
The planned network, as mapped in OpenStreetMap and clipped to the study area, is about 88 km: Línea 1 (under construction, opening 2028), its extension to Calle 100, Línea 2 (in tender), Línea 3 (proposed, Soacha to Calle 100) and Regiotram de Occidente. Turn on Same km as the plan to see the mould's 88 strongest kilometres.
To compare networks of the same length, I used a deliberately simple travel model. Without rail, every trip is a bus ride at 18 km/h along a route 30 % longer than the straight line. With rail, a trip can take a bus to the nearest rail point, wait 2.5 minutes on average, and ride at 35 km/h. Each trip takes the faster option, and trips are weighted by how many people live or work at each end.
| Network (88 km) | Average door-to-door time |
|---|---|
| Buses only | 50.5 min |
| Planned metro + Regiotram | 41.8 min |
| Physarum, no demand weighting | 39.3 min |
| Physarum, with demand (3 seeds) | 38.1 – 38.9 min |
With the same kilometres, the mould's network saves about three more minutes per average trip than the plan. 41 % of the planned kilometres run within 750 m of the mould's network.
The biggest disagreement is the shape. The plan is mostly radial: lines run into the centre and along Avenida Caracas and Calle 72. The mould keeps a strong centre too, but puts one of its strongest corridors on the diagonal from Suba to Soacha through the western localidades, where most people live.
Compared with TransMilenio
Bogotá already has a mass-transit network: TransMilenio's trunk busways, 109 km of them inside the study area. Turn on TransMilenio in the map to see them. To compare shapes rather than modes, I gave both networks the same speed and the same 2.5-minute wait.
| Network (109 km, same speed) | Average door-to-door time |
|---|---|
| TransMilenio trunk lines | 39.7 min |
| Physarum (3 seeds) | 36.7 – 37.5 min |
At the same length, the mould's layout is about two to three minutes faster per average trip, and 46 % of TransMilenio's kilometres lie within 750 m of the mould's network. TransMilenio follows the big avenues (Caracas, Autopista Norte, Calle 80, Américas, NQS, Calle 26), most of which converge on the centre. The mould uses fewer radial lines and adds the western diagonal.
My proposal
If I had to turn the mould's answer into a few lines, keeping Línea 1 since it is already being built:
- A western orbital, Suba – Engativá – Fontibón – Kennedy – Bosa – Soacha. It is the strongest corridor the mould builds that the plan does not cover.
- Regiotram del Norte to Chía, which the mould builds with no information about the proposal.
- A southern connection from Ciudad Bolívar and Usme to the centre, beyond today's TransMiCable.
What this is and is not
This is what a slime mold would build, not an engineering study. The model knows population and a few job centres; it knows nothing about real trip patterns, cost, geology, the existing TransMilenio network or property. The 30 % weight on job centres is my own assumption. The travel model ignores transfers and capacity. Treat it as a second opinion from an organism that is very good at networks.
What Colombia already runs
Colombia has one metro, in Medellín, and a long list of bus rapid transit (BRT) systems that took the metro's role in other cities.
| System | City | Type | Opened | Size |
|---|---|---|---|---|
| Metro de Medellín | Medellín | Metro (Lines A and B), plus cable cars and a tram | 1995 | More than 900,000 passengers a day (2025) |
| TransMilenio | Bogotá | BRT | 2000 | 114.4 km of trunk busway, 11 trunk lines |
| MIO | Cali | BRT | 2009 | 39.6 km of trunk lines (2019) |
| Metrolínea | Bucaramanga | BRT | 2009 | Passengers down 99.8 % in a year (DANE, Q4 2025) |
| Transmetro | Barranquilla | BRT | 2010 | 14 km, 2 trunk lines |
| Megabús | Pereira | BRT | 2006 | 16.7 km (2011) |
| Metro Línea 1 | Bogotá | Elevated metro | March 2028 (planned) | 23.96 km, 16 stations |
| Regiotram de Occidente | Bogotá – Facatativá | Regional tram-train | 2027 first stretch, about 2030 in full (planned) | 39.6 km, 17 stations |
I found no published Physarum study of a Colombian city or of Colombia's transport network, in English or Spanish. The nearest ones cover Mexico, Toronto and Argentina.
Medellín: the Tokyo test
Bogotá's comparison is a prediction: its metro does not exist yet. Medellín's does, so it allows the test Tero ran in Tokyo: give the mould the cities and see whether it rebuilds the network people actually built.
I ran the same model with the same parameters over the Aburrá Valley: Medellín's 16 urban comunas (2023 projections from the Alcaldía), Bello, Itagüí, Envigado, Sabaneta, La Estrella and Copacabana (DANE 2023), and five job centres (Centro, El Poblado, Olaya Herrera, the Central Mayorista and the Universidad de Antioquia). The comparison is metro Lines A and B, 30.6 km as mapped in OpenStreetMap.
The mould follows the valley north to south, as Line A does, but splits into two parallel strands: one west of the river through Robledo, Belén and Guayabal, and one east through the centre, El Poblado and Envigado. Line A runs along the river between them. Measured like Tokyo, it is again in the same range: cost 1.71 – 1.80, efficiency 0.83 – 0.86, and 16 % to 27 % of its length cuts something off if it breaks.
| Network (30.6 km) | Average door-to-door time |
|---|---|
| Buses only | 37.2 min |
| Physarum (3 seeds) | 32.0 – 33.2 min |
| Metro de Medellín, Lines A and B | 30.5 min |
Here the real network wins. In a narrow valley, one fast line down the middle is hard to beat, probably because, with only 30 km to spend, the mould's two strands each get only part of it. 41 % of the metro's kilometres lie within 750 m of the mould's network. That is the useful lesson for Bogotá: the mould is a good sketch of where demand pulls, not a replacement for engineering judgement, and in a city shaped like a corridor, the corridor line is the right answer.
Colombia, later
The same model can run over the whole country with cities as food and the Andes as light, to ask where intercity rail should go. That is the next experiment.
Sources
- Tero, A. et al. (2010). Rules for Biologically Inspired Adaptive Network Design. Science 327, 439–442.
- Adamatzky, A. (ed.) (2012). Bioevaluation of World Transport Networks. World Scientific.
- Map data, boundaries, land use and rail lines: © OpenStreetMap contributors (ODbL).
- Population: localidades from Wikipedia's Anexo:Localidades de Bogotá (El Espectador, 2020); municipalities from DANE 2023 projections.
- Empresa Metro de Bogotá for project status.
