Lecture 11 — Bayesian Optimization · AI4Chemical Sciences Bootcamp

Recording

Recording will be available after the bootcamp.

August 2026

Learning Objectives

Construct a Gaussian Process surrogate model and interpret its posterior mean and variance
Derive and implement Expected Improvement (EI), Upper Confidence Bound (UCB), and Probability of Improvement (PI)
Run a Bayesian optimisation loop over a discrete molecular library using BoTorch
Apply multi-objective BO (Pareto front) to simultaneous optimisation of potency and selectivity

Key Takeaways

Takeaway 1. The Gaussian Process posterior provides both a prediction and an uncertainty estimate in a single model — this joint output is what makes GPs the default surrogate for BO.
Takeaway 2. The acquisition function balances exploration (high uncertainty) and exploitation (high predicted value). UCB does this explicitly via a β parameter; EI integrates over the improvement distribution.
Takeaway 3. BO is sample-efficient because it conditions on all previous observations before choosing the next experiment — each new data point updates the surrogate globally, not just locally.
Takeaway 4. Standard BO scales poorly beyond ~20 dimensions. For high-dimensional molecular spaces, use latent-space BO (optimise in the encoder space of a VAE or GNN) or trust-region methods (TuRBO).