Data-Driven Optimization of an Industrial Batch Polymerization Process Using the Design of Dynamic Experiments Methodology

The optimization of batch processes usually relies on the availability of a detailed knowledge-driven model. However, because of the great varieties of industrial batch processes and their small production rates, a knowledge-driven model might not always be available. In such a case, a data-driven model, developed after a limited number of experiments, is an attractive alternative. Here we apply, in an evolutionary manner, the design of dynamic experiments (DoDE) (Georgakis et al. Ind. Eng. Chem. Res. 2013, 52 (35), 12369) methodology to model the process behavior and minimize the batch cycle time of an industrial polymerization process. In evolutionary DoDE, the initial design is selected conservatively in the close vicinity of the previous operating conditions to minimize the risk of violating safety constraints of the industrial process. After the initial data-driven model has been estimated using the collected data, an optimal operating condition satisfying process constraints is calculated. In addition, the input domain is enlarged to seek conditions that further optimize the process. The above steps are iterated until the most optimal process performance is achieved. We examine this evolutionary DoDE approach in silico using a detailed simulation of a working polymerization process at Dow to produce that data. After three rounds of experiments are performed, a 17.2% reduction in batch cycle time is achieved while all constrains on safety and product quality are met. It is only 0.7% longer than the batch cycle time obtained using model-based optimization, assuming a 100% accurate model is available.