Pareto最適実験：走査型プローブ顕微鏡における人間主導型多目的ベイズ最適化

Automated experimentation has the potential to accelerate scientific discovery across disciplines, but its success requires systematic methods to optimize multiple, often conflicting, objectives under uncertainty. We present multi-objective Bayesian optimization (MOBO) as a general framework for balancing competing rewards and integration of human guidance in autonomous experimentation. Rather than identifying a single optimum, MOBO constructs the Pareto front, providing a principled description of all trade-off solutions and revealing the interdependencies between partially known reward functions. This enables systematic exploration of parameter space and quantifiable decision-making. Importantly, MOBO naturally supports human-in-the-loop control. Researchers can reweight objectives or adjust reference points to steer experiments toward desired outcomes, incorporating expert judgment without breaking automation. Together, MOBO and human guidance transform experimental optimization from trial-and-error tuning into a reproducible, interpretable, and customizable process, offering a scalable methodology for building trustworthy and efficient self-driving laboratories.