Data-Driven Control Strategies for the Autonomous Operation of the Pharmaceutical Crystallization Process

In this contribution, we studied the deep neural network (DNN) for the control of the cooling crystallization of a model compound system. To this end, firstly the performance of the different neural network architectures in conjunction with the various combination of the time-series process data was tested for the training of the data-based model in order to assess the best fit model-training data architecture. The identified network model, which was trained with the offline process data, was utilized in a predictive control strategy. The objective of the control strategy was to optimize the supersaturation generating/decaying variable in the crystallizer to achieve a target crystal-state property profile throughout the process. The performance of the proposed control strategy was tested in the presence of the process disturbance and benchmarked against a radial basis function (RBF) based control strategy. The results showed that the DNN model was able to approximate the crystallization process input-output relation with R2 ranging between 0.767 and 0.990 and achieve the target profile at the end of the operation with a 22.3 % offset.