Finding optimal operating conditions fast with a scarce budget of experimental runs is a key problem to speeding up the development of innovative products and processes. Modeling for optimization is proposed as a systematic approach to bias data gathering for iterative policy improvement through experimental design using first-principles models. Designing dynamic experiments that are optimally informative in order to reduce the uncertainty about the optimal operating conditions is addressed by integrating policy iteration based on the Hamilton-Jacobi-Bellman optimality equation with global sensitivity analysis. A conceptual framework for run-to-run convergence of a model-based policy iteration algorithm is proposed. Results obtained in the fed-batch fermentation of penicillin G are presented. The well-known Baijpai and Reuss bioreactor model validated with industrial data is used to increase on a run-to-run basis the amount of penicillin obtained by input policy optimization and selective (re)estimation of relevant model parameters. A remarkable improvement in productivity can be gain using a simple policy structure after only two modeling runs despite initial modeling uncertainty.