Bayesian Optimization for Scientific Experiment Design

Bayesian Optimization optimizes expensive experiments by building a probabilistic response model and selecting the next experiment with an explicit decision rule.

AcquiLAB combines a surrogate model, an acquisition function, and an iterative experiment loop.

The loop continues until a target, plateau, budget, or iteration stopping condition is reached.

Gaussian Process surrogates estimate the unknown response surface and quantify uncertainty when data is scarce.

• Models nonlinear response surfaces from limited runs.
• Estimates uncertainty to guide high-value experiments.
• Updates after every accepted observation.

The acquisition function balances expected gain against uncertainty.

These modes match the current Product optimization catalogue. Parallel BO is available as Experimental batch behavior on the BoTorch BO strategy. Multi-fidelity, Active Learning, Physics-aware, and Mechanism-aware remain unimplemented and are not advertised as current Product modes. A separate strategy identity named parallel_bo is not part of the catalogue.

MOBO is used when objectives compete. It searches for Pareto-optimal trade-offs instead of claiming one universal optimum.

Exploration samples uncertain regions to improve knowledge. Exploitation samples strong predicted regions to improve objective values.

Typical stopping criteria include response plateau, objective threshold, iteration limit, or exhausted experiment budget. Stabilization does not guarantee a global optimum.