A liquid handler dispenses a fresh reaction mixture — reagent concentration and temperature set by the optimizer — into the plate, a flow reactor carries it through a continuously-flowing tube instead of a batch flask, and an in-line analyzer reports the reaction yield the moment it exits. That reading feeds back into a surrogate model of the whole condition space (the floating response surface): mean yield is estimated everywhere by distance-weighting the samples collected so far, and an acquisition score blends that estimate with how unexplored each region still is.
acquisition(x) = mean̂(x) + κ·distance_to_nearest_sample(x)
Bayesian mode always dispenses next at the highest-scoring point — balancing exploiting promising conditions against exploring gaps in the map — so it typically locks onto the true optimum in a handful of runs. Random search picks blindly and needs far more evaluations to get lucky; a full grid sweep of this same 36-cell plate would need 36 runs. Watch the response surface rise toward its true peak, and the gold marker track the best recipe found.