Berkeley Wildfire Spread and Access Explorer
Explore how different wildfire scenarios could affect Berkeley's neighborhoods and routes, and see which access vulnerabilities and possible mitigations recur across scenarios. Live at https://ericdf.github.io/berkeley-wildfire-evacuation-explorer/ (the repository keeps its name, berkeley-wildfire-evacuation-explorer).
A browser scenario explorer on Berkeley's real street network and terrain. Pick a fire, run it,
then see five results for that fire (each summarized over 15 runs, with a plain label saying whether it
holds in nearly every run, depends on fire strength, or varies from run to run) and which tested changes improve them.
Official instructions: the City's
Fire Weather and Evacuation page.
Why no evacuation times. The traffic model clears roads much faster than the City's 2025 Evacuation Time Study: with the fire off, its 90%-out times run 4–39% short of the study's on 7 of 8 calibration cases, more than 10% short on 3 (R02, the 1923 repeat, weekday 39% and night 12%; R05, Fire Zones 2–3, weekday 20%), and within 10% on Panoramic Hill (SOURCE_NOTES CD1–CD19, docs/model-fix-plan.md). So the explorer does not headline, rank or summarize with clearance times. Where a clock still appears (the run's clock, Expert view's results and Compare) it is labeled model time, not a forecast. The comparison with the study sits in Assumptions & limits, Model performance, as a known limitation. Product brief: docs/product-brief-spread-access.md (Eric, 2026-11-06).
The design brief is specification.md. Every insight taken from a document (the City's 2025 Evacuation Time Study, SB 99 and AB 747 reports, StoryMap and others), with page references and what we did with it, is in SOURCE_NOTES.md. What is still missing is listed in DATA_INVENTORY.md; problems found in other publishers' data, and whether they have been told, in ISSUES.md; where the project stands and what comes next, in HANDOFF.md. Design notes for work in progress are in docs/design/, and reports from finished analyses in docs/.
Where it stands (2026-11-04). The spec's Phase 1 (map, traffic and an illustrative fire on Berkeley's streets) is built, and most of Phase 2: interventions with costs, comparisons over many runs, and the Guided view's three tabs. Phase 3 (evidence-backed Berkeley inputs) is under way (docs/design/phase-3.md): City building footprints give the structures the fire reaches and an assumption range for those lost; LANDFIRE fuels are an opt-in setting; each zone's ACS persons per household and share of households without a vehicle are the default; enforcement of posted no-parking signs and daylighting is a plan setting. Structures do not yet change how the fire spreads, and home hardening is not modeled (docs/design/structure-spread.md).
Run it
No build step or dependencies. The simulator is plain ES modules and needs a local web server
(browsers block fetch from file://):
cd berkeley-wildfire-evacuation-explorer
python3 tools/serve.py
# open http://127.0.0.1:8766
Use 127.0.0.1 rather than localhost: if another server is listening on the same port
over IPv6, localhost may reach that one instead.
Loading takes a few seconds: about 10 MB of terrain and 2 MB of street network.
Playing
Guided view fits a phone: the map screen takes one screen with no page scrolling; drag the map with one finger, pinch to zoom, tap a street for its card. Expert view works on a phone but is built for a larger screen; on a phone it is one column: the map, then the choices, then the results.
Guided view (the default; rebuilt phone first on 2026-11-10, js/guided.js, js/guided-data.js, SOURCE_NOTES GV1–GV10). Four screens, the same on a phone and a desktop:
- Map. Pick one of six fires (a dry east wind from Grizzly Peak Blvd, the 1923 fire repeated today, Panoramic Hill, a fire from Tilden into the north hills, fires in buildings after a Hayward Fault earthquake, and a fire low in Claremont Canyon in an onshore west wind) and press Run simulation: one run of today's plan (first order 10 minutes after ignition, today's parking, a fall night, medium severity, seed 1), fire and traffic only. Afterwards, chips switch the map between Fire (streets colored by how often embers reached them in 15 runs), Ways out (streets with one modeled way out, colored by how often the fire front reached it) and Engine reach (engine drive time at ignition); tapping a street opens a short card.
- Results. Five cards, most important first: people not yet out when embers reached their street, when the fire front did, hill residents with no engine route within 5 minutes, hill residents on streets with only one way out, and cars left with no open route. Each figure is the median of today's plan's 15 runs of that fire (low, medium and high severity × seeds 1–5) in
data/model/mitigations.json, or the network figure in data/model/spread-access.json, with a plain label ("In nearly every run", "Depends on fire strength", "Varies from run to run"; the rules below) and an Explain this pop-up (what it counts, the runs, the provenance category, limits, the watched run's figure).
- Changes. Every tested change, grouped by the result it improves on that fire, derived from the table's verdicts (better in every run, or in three of four and never worse); the rest under "No consistent benefit in our tests". Each row gives the gain (on this fire, or on any fire for engine reach), its label and its cost; each group shows its top 3, with "Show N more" for the rest. Pick up to three and Run. Any difference counts as better or worse; the label says whether it holds across the runs.
- Comparison. Each result for today's plan and with the changes, the difference and its label: one change from the stored table, two or three from 15 paired runs in the browser.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
