1: \begin{abstract}

2: 	Experimental (design) optimization is a key driver in

3: 	designing and discovering new products and processes. Bayesian Optimization

4: 	(BO) is an effective tool for optimizing expensive and black-box experimental

5: 	design processes. While Bayesian optimization is a principled data-driven

6: 	approach to experimental optimization, it learns everything from scratch

7: 	and could greatly benefit from the expertise of its human (domain)

8: 	experts who often reason about systems at different abstraction levels

9: 	using physical properties that are not necessarily directly measured

10: 	(or measurable). In this paper, we propose a human-AI collaborative

11: 	Bayesian framework to incorporate expert preferences about unmeasured

12: 	abstract properties into the surrogate modeling to further boost the

13: 	performance of BO. We provide an efficient strategy that can also

14: 	handle any incorrect/misleading expert bias in preferential judgments.

15: 	We discuss the convergence behavior of our proposed framework. Our

16: 	experimental results involving synthetic functions and real-world

17: 	datasets show the superiority of our method against the baselines.

18: \end{abstract}

19: