WeeklyDynamo: Generative Design and Dynamo for AEC

GD–AI Loop 2.0 · Lab Notes #7 · From testbed to practice The previous six posts were notes from a simplified 1.5 m grid testbed. This one asks: “So what is it good for in actual design practice?” What was confirmed in the testbed and how it might carry over to practice (not yet validated) are separated by color. We then fed a single brief — “10 rooms, 2 floors” — into a fictional irregular lot and ran it end to end as one flow: from the instant answer, to a GA warm-started from past designs, to a comparison of egress evaluation criteria. After seeing those results, we added one wider lot, with room to choose exit stairs, as an example. Nov 1, 2026 (English edition Nov 6) · Previous: Lab Notes #6 — evacuation simulation as the fitness function Update, Nov 6, 2026. Takeaway ③ and §5 compare evacuation criteria with the v1 crowd model of Lab Notes #6, which has since been replaced by a validated model (Lab Notes #8, rewritten: continuous agents, SFPE door and stair limits, IMO test c...

GD–AI Loop 2.0 · Lab Notes #8 (rewritten) · validated crowd model → 20,947 buildings → AI The first version of this post trained a neural network on a crowd model that a fire engineer would call a toy: one person per 1.5 m cell, so the densest crowd it could form was 0.44 persons/m²; doors one cell wide; single-file stairs; no validation; three example buildings. This rewrite replaces the people model and shows the checks a practitioner would ask for before trusting any number — the IMO MSC.1/Circ.1533 test cases, SFPE flows and speed–density, and the same model run in JuPedSim. It then re-runs all 20,947 designs with occupant loads, pre-movement times and blocked-exit and blocked-stair scenarios, stages 120 of them at once in NVIDIA Omniverse, re-measures the Lab Notes #6 result, and retrains the network. Nov 6, 2026 — rewritten the same day after a practitioner's critique of v1 ("too elementary; why only two or three cases?") · Previous: Lab Notes #6 — evacuation sim...

GD–AI Loop 2.0 · Lab Notes #6 In the previous post our GPU generative-design pipeline scored evacuation with a travel-distance formula. This time the genetic algorithm's fitness function runs a crowd evacuation simulation — dozens of people queuing and slowing down in the exit stairs — as an NVIDIA Warp kernel on the GPU. The buildings it picks changed. Then we varied the site shape 320 ways and tested again whether machine learning can propose designs for unseen sites. Oct 25, 2026 (updated Nov 1–2; English edition and CPU vs GPU benchmark Nov 6) · Previous: Lab Notes #5, "80 million buildings in 42 minutes" Update, Nov 6, 2026 — the crowd model in this post has been replaced. The experiments below used the v1 crowd model: one person per 1.5 m cell (so at most 0.44 persons/m²), one-cell doors, single-file stairs, 42–62 people per building and no validation. Lab Notes #8 (rewritten) replaces it with a continuous agent model (the collision-free speed model used by JuP...