Variability of the composition and properties of raw materials used for biodiesel production may cause a loss of productivity, since the same operating conditions give rise to different yields for alternative feedstock sources. The capability to re-optimize the process when the raw materials change may lead to a significant improvement in productivity. For yield optimization, first-principles models of a biodiesel reactor have limited prediction capabilities due to the complex kinetics involving transesterification and saponification reactions, which demands active learning of relevant data through optimal design of experiments. In this work, a Bayesian approach for integrating experimentation with imperfect models is proposed to optimize biodiesel production on a run-to-run basis. Parameter distributions in a probabilistic tendency model for the transesterification of triglycerides are re-estimated using data from a sequence of experiments designed to guide policy improvement. Global sensitivity analysis is used to formulate the optimal sampling strategy in each dynamic experiment as an optimization problem. Results obtained highlight that, even when there are significant errors in the tendency model structure and reduced information content in samples, a significant increase in biodiesel production can be achieved after a handful of runs.