Surrogate models and acquisition functions — closing the loop between prediction and experiment for efficient molecular discovery.
Takeaway 1. At fixed GP hyperparameters, the posterior variance depends only on where you sampled, not on what you measured — so you can plan the informativeness of a campaign before running a single experiment.
Takeaway 2. The choice of acquisition function (EI, UCB, Thompson sampling) matters less than you'd think — on the EDBO reaction benchmark they converge to similar performance. Getting the noise model right matters far more: setting σₙ² ≈ 0 lets the surrogate interpolate a lucky measurement and turns the rest of the campaign into a lottery.
Takeaway 3. Categorical reaction variables (ligand, solvent, base) are chemically related to each other — one-hot encoding throws that away. Representation and kernel choice can move the needle more than the acquisition function does.
Takeaway 4. DOE and BO are complementary, not competing: DOE explains a system through main effects and interactions from a design fixed up front; BO adaptively hunts for the optimum. A common workflow screens with DOE to cut down the factor space, then hands the survivors to BO.