Reduced order models provide an efficient alternative to the rigorous models and are based on simplifying assumptions. In general, the accuracy of a model increases with increasing level of model complexity. However, the improved accuracy is often at the expense of the computational efficiency. The rigor of the models used in practice is therefore governed by a compromise between accuracy and computational efficiency. This paper introduces a novel concept of "Optimal Model Reduction" where the trade-offs between accuracy and efficiency are decided using nonlinear optimization techniques. The effectiveness of this new approach is presented in the context of batch distillation, for which a hierarchy of models exist in the literature.