In this article, we demonstrate a model-based approach for controlling the average size of crystals produced by batch cooling crystallization. The method is distinguished most notably in the modeling strategy. Rather than developing a crystallization model within the population-balance framework, as is usually done, we apply a machine-learning technique to identify an empirical model from measurement data. The model is low-dimensional and can therefore be discretized and used with dynamic programming to obtain optimal control policies for producing crystals of targeted average sizes in prespecified batch run times. Experimental results are reported that demonstrate the use of the identified policies to produce crystals of the desired average sizes in the specified run times.