The translucent gold hills are the true, unknown reaction-yield
surface over temperature and catalyst loading — a real lab never
sees this shape, only the points it samples (the white pins). The
wireframe surface is the optimizer's current belief, built from
a kernel-weighted average of samples seen so far. Each iteration it
picks the next experiment where the "acquisition function" — belief
plus uncertainty scaled by β — is highest, then samples there and
updates its belief. This is a simplified stand-in for Gaussian-process
Bayesian optimization used in real autonomous/self-driving labs.
acquisition(x,z) = mean(x,z) + β · uncertainty(x,z)
mean(x,z) = Σ wᵢ·yᵢ / Σ wᵢ, wᵢ = exp(−‖(x,z)−(xᵢ,zᵢ)‖² / 2ℓ²)
uncertainty(x,z) = 1 − Σwᵢ / (Σwᵢ + 1) (high far from any sample)
next point = argmax acquisition(x,z)
- Exploration weight (β) — low β exploits the best-known region; high β chases unexplored, high-uncertainty ground first.
- Auto-run speed — how fast the optimizer proposes and runs new experiments.
- Bayesian vs Random search — compare the guided acquisition strategy against picking the next point uniformly at random.