Scroll to begin
01 · The question
Research question
Can an AI lab design a cheap coating of five layers or fewer that beats Stanford's seven-layer cooler, using fewer simulations than standard search?
Stanford, 2014
Layers7MaterialsHfO₂ · SiO₂Solar reflected97%Cooling power40.1 W/m²Below air temp4.9 °C
Phys.io goal
Layers≤ 5MaterialsSiO₂ Al₂O₃ Si₃N₄ TiO₂ MgF₂No hafnium✓Target≥ 50 W/m²Budget400 sims
The bottleneck. Every candidate coating costs one physics simulation. Random search needs 184 on average to hit the target; the best standard optimiser needs 37.5. The lab wins only if it needs fewer.
02 · Control first
Trust the simulator before trusting a design
01Build Stanford's published stack7 alternating SiO₂ / HfO₂ layers on silver, thicknesses from Raman et al., Fig. 1dDone
02Simulate it on our optics engineTransfer-matrix optics, AM1.5 sun, clear-sky atmosphere, 300 KDone
03Solar reflectance 97.7% vs 97% publishedInside the 2-point tolerancePass
04Cooling power 11.8 vs 40.1 W/m² publishedOur analytic sky is harsher than Stanford's rooftop. Disclosed, so we compare simulated with simulated onlyGap shown
05Logged as an established factWritten to the shared research record; the design budget unlocksUnlocked
Policy control_first. No agent may simulate a new design until the control passes. Enforced by tool permissions, not by prompt.
50W/m²
One locked target · one simulator · one 400-simulation budget · for every method
The rule
Fewest simulations wins
03 · One discovery loop, six decisions
Each department owns one decision. A supervisor routes the work, and a human approves anything consequential. Built on Omnigent.
Seven departments.
One loop.
One loop.
04 · Faster
Simulations needed to reach 50 W/m²
Average simulations to target · lower is faster · 95% interval
Share of runs that reached the target
05 · Thermal emissivity vs Stanford · simulated━ best cheap design ┅ Stanford 7-layer
Best cheap design so far
Stanford
Cheap 4-layer
Same simulator and sky for both. Found by the expert-heuristic baseline (seed 6), not yet by the agent lab; the agent's best design replaces it after its run. Simulated only: needs fabrication and outdoor measurement before any real-world claim.
Rigor first, then speed
A speed-up only counts if nobody could have gamed it. Every claim on this page is backed by a run record or a cited source.
Next experimentRun the agent lab on the same 20 seeds, target and budget, plus an ablation without shared memory. Then fabricate the best design and measure it outdoors against Stanford's.
06 · Live lab
Give the lab a problem. Watch it work.
The agents run in Omnigent on the lab machine. Every message, tool call, record and log below comes from the run itself; nothing is scripted.
Connecting to the lab…
Agents and activity
Departments · what each agent is doing
Click a department to filter the activity
Net cooling power of every simulated design · dashed: Stanford control in our simulator
Solar reflectance vs 8–13 µm emissivity · up and right is better
Report