Polyolefin Process Improvement Using Machine Learning

Data Analytics have emerged to play an important role in chemical process industry in recent years. With large amounts of data being generated in process plants with the wide use of distributed control systems, industries are looking to extract information from the data by using Data Analytics/Machine Learning methods. Process data analytics and process monitoring is particularly useful for polymer processes since the process measurements are more frequent and precise, while the output measurements for polymer processes like molecular weight are sparse and noisy. Thus, data-based sensors derived from machine learning and multivariate statistical methods for polymer quality measurements are critical for process monitoring and quality control. We can also use these techniques to identify process outliers and their cause of being an anomaly.

We focus on some novel aspects of the application in machine learning in polymer processes in this study which have not been studied extensively in prior literature. We showcase an inverse modeling approach to predict the operating conditions of the polyolefin process if we know the quality parameters of a polymer grade. We also explore the application of semi-supervised machine learning methods to improve the prediction of polymer quality parameters.

We consider different polyolefin processes producing Polyethylene and Polypropylene for our analysis. We also showcase how to simulate plant data using dynamic process models when plant data is not available. The process variables considered here are flow rates and reactor operating conditions while the product quality variables used are the polymer Molecular weights/ Melt Index, density and polymer production rate. We showcase the use of predictive models like ensemble based regressors for predicting product quality indicators. We also demonstrate the utility of causal models like partial least squares to study the causal effect of the process parameters on the polymer quality variables. We make use of anomaly detection methods as well to identify the process outliers and also the reasons for their outlier behavior.

We demonstrate an inverse modeling approach to predict the operating conditions for particular polyolefin grades given the product quality targets. If we want to produce a new polymer grade, we can predict its operating conditions using this approach. In the current changing market scenarios with different customers and end uses of polymer products, the requirement of polymer quality targets keeps on changing. Therefore, having a methodology to predict the process operating conditions for producing a new polymer grade becomes critical. This approach is particularly useful where causality of the variables is unknown.

The output measurements for polymer processes like molecular weight are sparse and noisy while there is lot of continuous process data available. Thus, semi supervised machine learning methods can be utilized to incorporate the vast amount of unlabeled data for data analysis. We used semi-supervised learning techniques like pseudo labeling for utilizing the unlabeled data along with the labeled data. The semi supervised learning techniques which includes the labeled as well as unlabeled data gave much higher prediction accuracy for polymer quality parameters than considering only the labeled data in traditional supervised learning methods.