Preference-Aware Multi-Objective Bayesian Optimization for AI-Guided Polymer Design

In polymer design, multiple conflicting properties must often be balanced under limited evaluation budgets. Multi-objective Bayesian optimization (MOBO) is promising for AI-guided design, but methods approximating the entire Pareto front may spend costly evaluations on regions irrelevant to user preferences. Furthermore, directly specifying trade-offs becomes difficult as the number of objectives increases. Preference weight vectors, whose components sum to one, encode relative importance across objectives and can focus MOBO on specific Pareto-front regions, but their correspondence to resulting trade-offs is unclear. We propose Preference-to-Pareto MOBO (P2P-MOBO), which maps preference weight vectors to estimated Pareto-front points in material property space. This enables preference adjustment through estimated property trade-offs while focusing optimization on preferred regions. Experiments on a polymer dataset show that P2P-MOBO generally identifies candidates closer to preferred targets than baselines under the same evaluation budgets.