RepMOBO: Representation-Guided Multi-Objective Bayesian Optimization for Pareto-Optimal MOF Discovery

Abstract Screening metal–organic frameworks (MOFs) across chemical space by simulation or experiment is costly, while trade-offs among adsorption objectives hinder Pareto-optimal discovery. Material representation learning offers a promising route by encoding crystal structures into compact feature spaces. However, prior work has been confined to offline prediction, leaving unclear how representation choice shapes search behaviour and Pareto discovery in multi-objective Bayesian optimization. Here we propose RepMOBO, a representation-guided multi-objective Bayesian optimization framework, using multi-vapour MOF adsorption to examine how handcrafted physicochemical descriptors and pretrained structural embeddings shape low-budget Pareto discovery. At a 10% query budget, the hybrid-representation configuration yielded 4.5-fold more configuration-specific strict Pareto-front discoveries than the handcrafted-descriptor configuration. RepMOBO-prioritized candidates guided experimental screening, yielding synthesizable MOFs with target-dependent, state-of-the-art adsorption performance. These results reveal a prediction–discovery mismatch: strong offline prediction need not translate into efficient low-budget Pareto discovery. RepMOBO therefore elevates representation from static encoding to an active search variable and motivates transferable representation–surrogate–policy co-design for costly multi-objective discovery in numerically representable candidate spaces.