The Mold That Mapped Tokyo: Physarum polycephalum and Biologically Inspired Network Design

In 2010, a team of researchers led by Atsushi Tero published a paper in Science that stopped the network-engineering world in its tracks. They had placed a plasmodial slime mold, Physarum polycephalum, on an agar plate representing the Kanto plain around Tokyo, with oat flakes marking the positions of 36 major cities. Within 26 hours, the mold had grown a web of nutrient-transport tubes connecting those food sources. The resulting network bore a striking resemblance to the actual Tokyo rail system - comparable in efficiency, fault tolerance, and total construction cost [1].

The finding was remarkable not because a slime mold is clever, but because it is not. It has no neurons, no memory in any conventional sense, and no representation of the map it inhabits. Yet through purely local chemical and mechanical interactions, it converged on a solution that a team of human engineers had spent decades refining. That gap between the simplicity of the mechanism and the quality of the outcome is the central puzzle - and the central lesson - of physarum network research.

A Single-Celled Architect

Physarum polycephalum belongs to a group called plasmodial slime molds. It is not a plant, not an animal, and not a fungus - it occupies its own kingdom, Amoebozoa, and in its feeding stage exists as a single giant cell containing many nuclei, all sharing one continuous cytoplasm. This structure, the plasmodium, can spread across many centimetres of substrate while remaining physiologically one entity.

The plasmodium moves and transports nutrients through cytoplasmic streaming: rhythmic contractions of the cell cortex drive cytoplasm back and forth through a branching network of tubular channels. These channels are not fixed - they widen when flow through them is high and narrow when flow is low, a feedback loop that continuously remodels the network in response to nutrient distribution. It is this adaptive remodelling that makes the organism useful to network scientists. Nakagaki et al. demonstrated the principle dramatically in 2000 by showing that a plasmodium placed at the entrance of a maze would, within hours, retract all branches that hit dead ends and concentrate its mass into a single tube running the shortest path to the food at the exit [2]. The cell shape and the contraction pattern are tightly coupled: spatial gradients in oscillation phase drive the directional streaming that redistributes mass [3].

The Tokyo Experiment

Tero and colleagues set up their experiment on a flat agar plate shaped to match the coastline and mountainous terrain of the Kanto region, with those areas covered in light (which Physarum avoids) to mimic geographical constraints. Oat flakes - a preferred food source - were placed at the positions of 36 population centres, with Tokyo at the centre of the arrangement. A small inoculum of plasmodium was placed on the Tokyo position, and the network was photographed at regular intervals over 26 hours [1].

The mold's network evolved through a recognisable sequence. First, the plasmodium spread outward in all directions, forming a dense exploratory mesh. Then, over subsequent hours, lightly used connections were pruned while heavily trafficked tubes thickened, leaving a sparse but robust skeleton. When the authors overlaid the final network on a schematic of the Tokaido rail system, the correspondence was striking: major trunk connections appeared in both, as did several redundant loops that provide alternative routes when individual links fail.

To make the comparison rigorous, the team quantified three properties of the mold network and the real rail network: total tube length (a proxy for construction cost), transport efficiency (average shortest-path length between city pairs), and fault tolerance (robustness to the removal of individual nodes). The mold's network matched or exceeded the engineered system on all three metrics, and it did so repeatedly across multiple experimental replicates [1].

The Mathematics of Slime

One of the contributions of the Tero paper was a mathematical model that reproduces the mold's behaviour from first principles. The network is represented as a graph: food sources are nodes, and the tubes connecting them are edges each carrying a conductivity D_ij and a physical length L_ij. Flow through each tube obeys Kirchhoff's pressure law - the volumetric flux Q_ij is proportional to the conductivity and to the pressure difference between the two endpoints:

Pressures at all nodes are found by enforcing conservation of flux (Kirchhoff's current law), with one node designated as the source and another as the sink. The crucial dynamics come from the conductivity update rule: each tube's conductivity grows in response to the magnitude of the flux through it and decays in the absence of flow:

Here mu is a decay constant and f is a nonlinear reinforcement function. Tero et al. used f(Q) = |Q|^gamma, where the exponent gamma controls the sharpness of the feedback [1]. When gamma is above zero, tubes with slightly higher flux receive disproportionately more reinforcement than weaker tubes, amplifying small differences until the network self-organises into a configuration that efficiently routes the required flow. This is a positive feedback loop operating on a continuous variable - analogous in spirit to pheromone reinforcement in ant colonies, but acting directly on the physical conductance of each channel.

Why It Works

The power of the model lies in the tension between two opposing pressures. The decay term (-mu * D_ij) continuously prunes every tube, favouring networks with fewer and shorter connections - low construction cost. The reinforcement term f(|Q|) protects tubes that carry high flux, which tends to favour redundant paths that distribute load and maintain connectivity when a node is removed - high fault tolerance. The organism settles into a network that is near-optimal on both criteria simultaneously [1].

The gamma parameter acts as a dial between two extremes. At low gamma values the feedback is nearly linear, differences are barely amplified, and the resulting network retains most of its initial connections - a dense, mesh-like topology with high redundancy but high cost. At high gamma values the feedback is sharply nonlinear, weak tubes are eliminated aggressively, and the network converges toward a minimum spanning tree - cheap, but fragile. The real Tokyo rail system corresponds to an intermediate gamma, and so does the slime mold's spontaneous solution. Nature, through evolution, has landed on a parameter regime that human engineers arrived at by deliberate design.

From Petri Dish to Algorithm

The Physarum model has since been abstracted into a general-purpose network optimisation algorithm. Researchers have applied it to the design of road networks, power distribution grids, and internet routing topologies - problems where the goal is always some version of the same three-way trade-off between cost, efficiency, and robustness. Adamatzky's extensive work on Physarum computing explored the organism itself as a substrate for unconventional computation, demonstrating logical operations and path-finding on physical agar plates [4].

Compared with other bio-inspired approaches, the Physarum algorithm occupies a distinct niche. Ant Colony Optimization builds solutions incrementally through discrete, probabilistic decisions guided by pheromone trails - well-suited to combinatorial problems such as routing and scheduling. Genetic algorithms work over a population of candidate solutions, recombining and mutating them across generations - better suited to high-dimensional parameter spaces. The Physarum algorithm, by contrast, operates on a continuous physical network, evolving its structure in parallel across all edges simultaneously. It is less flexible than ACO or genetic algorithms, but for problems that are naturally expressed as network flows, it is remarkably direct. You can see both approaches in action on the Simulations page, where an interactive Physarum network simulation sits alongside the ACO demonstration.

Perhaps the deepest lesson of the slime mold experiment is methodological. Before the Tero paper, network designers optimised for one objective at a time - cost, or efficiency, or robustness - and then checked what the solution did to the other two. The plasmodium does not know about these objectives. It only knows about local flux and local chemical concentration. That it nevertheless finds a Pareto-efficient solution - one that cannot be improved on any single dimension without worsening another - suggests that the right local rules can implicitly encode global optimality criteria. Finding those rules, in biology and in engineering, remains an open and rewarding problem.