Inspired by arts.cern, Atmospheric Collider is an agentic art installation. Each day at midnight UTC, it fetches solar weather data from NOAA's Space Weather Prediction Centre and translates it into a deterministic harmonograph mandala. Same day, same form. No two days alike.
Collision trace
K-index at 1.0 — the magnetosphere barely trembles, a whisper in the solar cavity. Proton flux of 0.38 pfu generates a minimal freq1 of 1.58 Hz, while the solar wind drifts at 367.1 km/s, too slow to disturb the geometric lattice. A four-fold symmetry emerges: the mandala of a quiet Tuesday.
Generated by claude-sonnet-4-5
At midnight UTC, a Vercel Cron job calls three unauthenticated NOAA SWPC endpoints — planetary K-index, integral proton flux, and solar-wind plasma speed — and distils them into a single storm classification.
The raw measurements are handed to an AI agent harness built on the Vercel AI SDK. A system prompt casts Claude as a Resident Scientist: it reasons about the atmospheric conditions and emits a structured JSON geometry specification — frequencies, symmetry order, colour temperature, and a chaos seed — all constrained by a Zod schema.
The geometry spec and the agent's full reasoning trace are written to a Supabase Postgres table with a UNIQUE constraint on the date column. Re-running the cron for the same day is a safe no-op — the existing trace is returned unchanged rather than overwritten.
The chaos seed drives a seeded PRNG (mulberry32) that seeds a dual-pendulum harmonograph. The resulting 15,000-point path is rendered as a Three.js line group with additive blending — same seed, same mandala, every time.