Substrate-Adaptive Decarboxylative Olefination Enabled by Active Transfer Learning

Generality in synthetic chemistry is rarely achieved through a single universal protocol. Instead, widely adopted transformations derive their utility from a well-characterized reaction space that enables rapid substrate-adaptive optimization. Here, we demonstrate how expert-guided reaction-space design combined with user-friendly machine-learning tools can accelerate the construction of such reaction spaces. Using the organocatalytic decarboxylative olefination of malonic acid derivatives as a model system, Bayesian optimization was used to generate a compact, transformation-specific dataset that served as transferable prior knowledge for substrate-adaptive optimization. This approach addresses key limitations of traditional fixed-condition substrate-scope studies by enabling rapid identification of high-yielding conditions for previously unseen and challenging coupling partners. Importantly, the learned reactivity landscape proved transferable beyond the substrate classes represented in the original reaction-development dataset, enabling efficient optimization of related nucleophile families. The resulting platform provides a mild, atom-economical, and highly stereoselective alternative to classical phosphonate-based olefinations, enabling access to diverse α,β-functionalized unsaturated thioesters, esters, and amides, including complex molecular architectures. More broadly, this study demonstrates that reaction-development datasets can serve as transferable reaction knowledge, enabling efficient substrate-adaptive optimization and accelerating the maturation of emerging synthetic methods into practically general transformations.