The atlas
86,622 nodes and 197,068 weighted links, each carrying a parameter from a public data set.
The atlas is the world energy system drawn as a graph and made to move. Every node holds one measured quantity and names where it came from. Apply a change anywhere and the effect propagates along the links until it settles, which takes between twenty and fifty steps.
The nine layers
| Layer | Nodes | Contents |
|---|---|---|
| Space | 19 | The measured solar constant, and the annual mean insolation above the atmosphere at each band of latitude, computed from orbital geometry |
| Weather | 2 | El Nino / La Nina and the North Atlantic Oscillation, from their measured index series |
| Climate | 2 | Carbon dioxide and the global temperature anomaly, both global series shown at Mauna Loa, where the record is kept |
| Events | 9,259 | Earthquakes of magnitude 5.5 and above since 1960 that reach infrastructure, and the earlier disaster set |
| Markets and fuel | 3,627 | Price benchmarks, national fuel supplies, and every port in the World Port Index |
| Grids | 3,546 | National grids and first-level administrative districts at their settlement centroid |
| Power plants | 34,936 | Every unit in the world database, 167 countries |
| Demand | 35,207 | Every settlement above fifteen thousand people |
| Behaviour | 24 | Behavioural channels, inside 840 drawn brain structures, fed by every district with a measured population |
Seven of the nine layers are drawn on a sphere. Mercator inflates Greenland to the size of Africa and stretches the high-latitude Russian and Canadian grids into ribbons, which is where a good deal of this system's fuel lives. On a globe every node sits at its true angular position and a point on the far side is hidden by the planet rather than flung to the edge of a rectangle.
The space layer is drawn standing off the surface, because the sun is not on the Earth and neither is the sunlight arriving above the atmosphere.
The last layer is not a place. Behaviour is drawn inside a brain, with 24 channels at anatomically defensible positions among 840 named structures of the Allen Mouse Brain Common Coordinate Framework. The structures are scenery: they carry no weight, no link and no channel, and they are held outside the node arrays so that nothing can mistake them for part of the model. The channels carry weight and wire into demand.
What a run does
Select a scenario or click any node and push on it. Each node takes the mean of what its neighbours are carrying, damped by its own resilience, bounded by a hyperbolic tangent so that nothing runs away, and the source is held at the value it was pushed to. The run repeats until the largest remaining movement falls below a threshold, which takes between twenty and fifty steps. Nodes the run moved pulse, with the depth of the pulse scaling with the size of the effect. Clicking a node draws its web of influence three hops deep.
The mean rather than the sum is the load-bearing choice. A national grid carries thousands of plants, and a node that added what its drivers were carrying would multiply a change by its own degree rather than distribute it. Each link is also divided by the number of links reaching that node from the same layer, so a layer speaks once however many members it has.
Most links carry in both directions: a grid outage reaches the plants that feed it as well as the cities below it. Links out of the climate and event layers carry one way only, because an earthquake acts on a grid and no grid causes an earthquake.
The atlas carries 60 prepared changes, grouped by kind. Each one names the number its size rests on, because turning a real event into a dimensionless push is where a model quietly becomes fiction. The rule is that a change of 1.0 is the largest of its kind in the record: the 1973 oil shock is 1.0 because crude roughly quadrupled in real terms and nothing since has beaten it, and everything else is measured against that.
Some magnitudes are derived rather than asserted. A country's fuel shares are already measured and sit in the model, so withdrawing its coal pushes its grid by the generation it actually loses; that number is different for every country and none of it is chosen. The El Nino scenarios are the observed peak of each event divided by the largest peak in the record. Where no number could be found, the scenario says so and calls itself an assumption.
| Category | Changes | Widest reach | Widest |
|---|---|---|---|
| People & cognition | 6 | 3,004 | Fear becomes salient |
| Universe & Earth, exogenous | 10 | 77,669 | A step in climate forcing |
| Country policy changes | 8 | 13,969 | The United States retires coal |
| Climate goal meetings | 8 | 76,165 | Paris, fully met |
| Natural disasters | 6 | 13,969 | A major California earthquake |
| Global pandemics | 6 | 2,744 | The Black Death, at today's scale |
| Wars | 5 | 5,478 | A war of 1939-45 scale |
| Technology improvement & buildout | 5 | 6,983 | Fusion arrives at scale |
| Energy makeup evolution | 6 | 5,121 | Renewables pass half of world power |
These are not forecasts. The model answers what is connected to what and how strongly, so a prepared change is a way of asking that question from a particular starting point.
Reach spans five orders of magnitude, and starting at the most connected power plant in the world moves one other node. That spread is the point. A model that answers the same number to every question is not answering.
What is above the climate
The sun is measured. The solar constant here is the NRLTSI2 composite, 1362.00 W/m² in 2023, with a stated uncertainty of 0.61. Across a solar cycle it moves by about one part in fifteen hundred, which is the size of the change the solar minimum scenario applies.
The insolation bands are not measured, and they are not estimated either. The annual mean sunlight arriving above the atmosphere at a given latitude is orbital geometry, and it is computed here from the measured solar constant, the obliquity and the eccentricity. The check that it is right is that the area-weighted global mean of the computed field has to equal the solar constant divided by four, the ratio of the disc the Earth intercepts to the sphere it radiates from. It does, to one part in ten thousand.
Each band reaches the national grids at its latitude, weighted by the share of that country's generating capacity that is solar, measured from the plant layer. Ninety-one countries have a solar fleet large enough to record. The rest get no link, because a country with no solar plants has no exposure to the sunlight for the model to carry.
Weather
Two modes of year-to-year variability carry their own measured index series. El Niño and La Niña enter as the Oceanic Niño Index, 917 overlapping seasons since 1950. The North Atlantic Oscillation enters as the standardised station-based index, 917 months over the same period.
The ENSO weight was not chosen. Regressing the global temperature anomaly on the annual ONI over 1950 to 2025 gives a correlation of 0.44, significant at better than one in a thousand, and a slope of 0.068 °C of global anomaly per degree of ONI. The correlation on the raw series is 0.10, because the temperature carries a strong trend and the ONI does not, so most of what a raw correlation measures is the trend. A quadratic in time is removed from the temperature first. The 0.44 is the link weight.
The NAO reaches two countries. Winter NAO and winter electricity demand in Great Britain are anti-correlated at −0.67, which is a published, quantified relationship, and there is an equivalent study for Ireland. A negative NAO winter is cold and calm: demand rises and the wind does not blow. That is the link. The NAO plainly affects more of Europe than two islands, but a weight is a number, and for the rest of the continent I did not have one I could cite, so those links are not drawn. The layer is deliberately smaller than the physics.
One correction worth stating, since it is the usual explanation for why Ireland is habitable at the latitude of Labrador. It is mostly not the Gulf Stream. Seager and colleagues took the ocean heat transport out of an atmospheric model and the winter contrast across the North Atlantic barely moved; the warmth is carried by the prevailing south-westerlies, by the seasonal release of heat the ocean absorbed in summer, and by the standing wave the Rocky Mountains impose on the flow. The ocean stores and releases; the atmosphere is what carries.
The behaviour layer
Behaviour is wired in both directions: the channels reach the consumer groups, and every district that resolves to a real administrative division reaches the channels.
Each of those 2,761 districts carries the population GeoNames records inside it and the concentration of that population, a Herfindahl index over its settlements: 1.0 means a single city holds everyone, and small means the population is spread across many towns. Kerala comes out at 0.012 and Bulawayo at 1.000, which is what those two places look like. Districts are joined to the settlement data on their administrative code, so the join is exact rather than a spatial guess.
The concentration is the weight of the link from a district into the behaviour layer. That the measurement is real does not make the direction real: that a more concentrated population couples more strongly to shared behaviour is a hypothesis, and it is marked assumed rather than measured.
The behaviour tab switches between the brain and the map, because the two things worth seeing are the channels and the places feeding them.
What is deliberately absent: there is no weighting of a district by the values of the people in it. The World Values Survey measures its two dimensions at national level for about 120 countries, subnational data exists for a handful, and nothing exists at the resolution of a state or a congressional district. Putting a value set on a district would mean spreading a national average across geography and calling the result a measurement. The concentration figure is a population statistic and says nothing about anyone's politics.
Connectivity
87 of the 86,622 nodes have no link to anything, which is 0.10 per cent. The figure is checked on every build, because an isolated node is a node the model cannot answer any question about.
An earthquake with nothing inside its felt radius is not added at all. Such a record is real and it affects nothing, and keeping it would make the layer look full while leaving most of it inert.
Reading a dense scene
86,622 points is more than anyone can take in at once, so the sidebar carries a threshold that hides everything below a chosen weight. Weight is generating capacity for a power plant and resilience for everything else, which is the quantity the propagation already uses. Hidden nodes still take part in a run. The threshold changes the view, not the answer.