Transfer learning for reaction optimisation: overview of its applications and limits

Reaction optimisation remains a critical yet resource-intensive step in chemical synthesis and process development, particularly in complex, multidimensional reaction spaces. Recent advances in automated experimentation and machine-learning-guided optimisation have significantly improved the efficiency of identifying optimal reaction conditions. However, most optimisation campaigns still operate in a cold-start regime, failing to exploit the growing body of historical experimental data. In this Review, we examine the emerging role of transfer learning as a strategy to overcome this limitation by enabling the reuse of knowledge from related reactions or prior optimisation campaigns. We provide a comprehensive overview of transfer learning methodologies applied to reaction optimisation, including multi-task Bayesian optimisation, ensemble approaches, neural-process-based models, hybrid artificial intelligence frameworks, and active transfer learning strategies. Through selected case studies, we highlight how these approaches can accelerate convergence, reduce experimental effort, and improve data efficiency, particularly in automated and flow chemistry platforms. We further discuss key challenges, including the risk of negative transfer, the difficulty of defining reaction similarity, and the need for chemically meaningful representations. Finally, we outline future directions toward knowledge-accumulating autonomous laboratories, where transfer learning enables cumulative chemical learning across reaction systems.