This tool couples a grid-based hazard model to a simplified road-network queueing model on a synthetic domain. It is designed to make the shape of evacuation risk legible — earlier/staged departure vs. later/simultaneous departure, and the effect of wind, rain, humidity, and road capacity — not to predict any real fire, storm, or town.
Fire model — known simplifications. Fire spread is a stylized hazard-growth model / evacuation-timing clock: neighbor-to-neighbor ignition probabilities shaped by wind (directional, elliptical), humidity, rain, and slope. It is not a numerical fire-behavior model, and a fire analyst should discount it accordingly on three specific points: (1) fuel is a single scalar value per cell, not a fuel-type/fuel-model system; (2) spread is neighbor-to-neighbor only, aside from the simplified ember-spotting effect described below there is no real firebrand-transport physics; (3) the rate-of-spread readout is a diagnostic sanity-check derived from the cellular automaton's own front advance, not an independent physical prediction.
Ember spotting (stylized). Each tick carries a small, wind-scaled random chance of igniting a new "spot fire" cell downwind of the active front, ahead of the contiguous burn area — a simplified stand-in for firebrand lofting/spotting, which in real wildland-urban-interface fires is a major cause of evacuation routes being cut faster than the visible front alone would suggest. It is a coarse stochastic effect, not an ember-transport model (no plume physics, spotting-distance distributions, or receptor fuel-bed ignition probability).
Hurricane model — known simplifications. Hurricane mode uses a growing radial wind/surge footprint advancing from a coastline, with low-elevation cells flooding earlier — a stylized analog of storm surge behavior. It is deliberately symmetric: it does not model storm-relative asymmetry, forward-speed-driven surge enhancement, or nearshore bathymetry, and it is not a SLOSH-class model.
Traffic uses a capacity/queue (cell-transmission-style) model per road segment, not individually simulated drivers.
Color and accessibility. The congestion ramp (green → yellow → red) and the hazard ramp (pale straw → orange → deep red) are not fully colorblind-safe in isolation. Every color-coded state is also backed by a text or numeric label — % jam, % safe, at-risk counts, "AT RISK"/"CUT OFF" zone flags, dashed line + X markers for blocked roads — so no reading depends on color perception alone.
What would make this operational: import real jurisdiction roads (GeoJSON/shapefile) and census population/housing data in place of the synthetic town; calibrate road capacities to HCM (Highway Capacity Manual) or a regional travel-demand model instead of illustrative constants; explicitly model no-vehicle and transit-dependent households and shadow (self-)evacuation behavior rather than one idealized queueing curve; and add scenario save/export so specific what-if runs can be archived and shared outside the browser session.