Batch Process Design Including Initial and Operating Conditions and Online Property Estimation in Acrylic Resin Polymerization

The objective in this study is to develop a system to achieve target ranges of product properties when acrylic resins are produced in a batch process. First, machine learning models between product properties and initial conditions are constructed with machine learning; a large number of candidates of initial conditions are generated; the candidates are input into the models to predict values of product properties; and the candidates whose predicted values are within the target ranges are proposed. Next, soft sensors are developed to estimate product properties in real time from spectral data obtained by near-infrared spectroscopy during batch process operation. Finally, machine learning models are constructed between operating conditions and product properties estimated by the developed soft sensors. The proposed system was used in an actual process to produce acrylic resin in a batch reactor, and it was confirmed that the batch reactor could be operated so that the product properties are within the target ranges, while product properties were continuously estimated in real time.