Learning Chemoselectivity in a Low-Data Regime: Bayesian Optimization of the NBS-Activated Selenocyclization of α-Allenols

The activation of diaryl diselenides with N-bromosuccinimide generates two competing electrophiles, and with α-allenols as the π-nucleophiles, product mixtures of brominated and selenylated dihydrofurans can arise. This selectivity challenge was quickly resolved through a Bayesian optimization campaign over seven reaction parameters covering >2000 discrete combinations. Fifteen randomly sampled experiments were sufficient for Gaussian-process surrogate models that, when asked to simultaneously maximize yield and selectivity, recommended conditions that delivered the 3-selanyl-2,5-dihydrofuran in up to 74% yield with complete suppression of the vinyl bromide. The identified protocol for the 5-endo-trig cyclization transferred without re-optimization to a range of structurally diverse α-allenols (50–81% isolated yield) and proved to be applicable to other dichalcogenides, as well as to the 6-endo-trig cyclization of related allene homologues.