← Back to projects

Physarum urban

research into slime mold capabilities with AI

I studied the slime mold Physarum polycephalum as an algorithm, testing it on real city maps with AI. Here is what I think holds up.

What it is

Physarum is a single cell with no brain. Tubes carrying flow thicken and unused ones shrink, leaving a short network with backup routes. In Tero et al. (Science, 2010) it linked oat flakes placed like towns around Tokyo into a network comparable to the region's rail system.

Two dark panels. Left: a thin road network connecting eight labelled hubs such as downtown, a hospital and a train station. Right: thick gold lines showing the backbone the slime-mold model kept.
A test town's eight hubs, and the backbone the model kept: 51% of the road length.

What it could do

  • Transit backbones. It shrinks a street map to a ranked backbone, and one setting slides it from lean to redundant.
  • Finding weak points. It ranks links by the flow they carry, which my ARIA project used to flag the bridges a supply network leans on; cut one and it reroutes.
  • Routing around obstacles. Pipes and wiring through crowded space are a possible use: a 2022 slime-mold-inspired method wasn't slowed by obstacles, though it wasn't tested head-to-head.
  • Growing parts. Fed at load points, it grows a frame along the load paths; my 3D-printable drone frame has only a rough hand calculation behind it so far.
  • Paired with AI. Claude turns a goal like "avoid flooding" into solver settings, then drafts the write-up.

What I built

  • In Davis, California, I cut 301 km of OpenStreetMap roads to a 105 km lean or a 167.5 km redundant backbone.
  • A flood-aware bike plan for Somerville and Medford, Massachusetts. Closing every street that touches the 100-year floodplain leaves about 28% of homes with no route to any Green Line Extension station, mostly because it closes the Mystic River bridges.
Street map of Davis with ten transit hubs and routes coloured by service tier, from thick dark red for every 10 minutes to thin brown for hourly feeders, with a legend and a 1 km scale bar.
Davis: the lean backbone's flows sorted into four bus tiers, from every 10 minutes to hourly (fixed cut-offs, not a ridership model). Map data © OpenStreetMap contributors.

The catch

I found no verifiable case of a transport agency adopting a Physarum design, and the Tokyo result shows two optimisers agreeing, not biology beating engineers. It skips most of real planning, its fault tolerance has only been tested against random breaks, I haven't compared it head-to-head with a standard Steiner-tree method, and the AI memo once mixed up two of my numbers.