Abstract
<jats:p>Transfer learning offers a promising route to accelerate Bayesian optimization (BO) by reusing knowledge acquired during related optimization tasks. In chemical reaction optimization, this most naturally arises during search-space expansion, where previous optimization campaigns involving related–but distinct–substrates, ligands, catalysts, or optimization objectives create opportunities to accelerate the exploration of new chemical systems. Despite the practical importance of this setting, BO-based transfer learning for chemically relevant search-space expansion has received comparatively little attention. Here, we systematically investigate the factors governing successful transfer learning using representative high-throughput reaction optimization benchmarks. We show that robust transfer learning depends not only on the transfer strategy itself, but also on the underlying Gaussian Process (GP) model. In particular, well-calibrated, dimension-aware hyperpriors substantially improve transfer performance over legacy default priors, while hidden-space molecular representations become increasingly advantageous as optimization campaigns are transferred to larger sets of previously unseen chemical species. More importantly, we highlight that the central challenge of transfer learning lies in balancing positive and negative transfer. Across the benchmark datasets considered here, direct reuse of historical observations proves remarkably effective, whereas deliberately constructed stress-test scenarios reveal that misleading prior information can substantially impair optimization performance when source and target campaigns differ sufficiently. Motivated by these observations, we introduce a simple two-phase transfer protocol that initially exploits historical observations before transitioning to task-aware transfer once sufficient target-task evidence has been collected. Although no single strategy is universally optimal, our results establish practical guidelines for selecting transfer-learning strategies according to the expected similarity between historical and target optimization campaigns, while identifying the proposed two-phase protocol as a robust default when this similarity is uncertain.</jats:p>