AIChE 2022 Spring Meeting

With careful consideration to the planning, capture, structure, evaluation and preprocessing of an experimental dataset, data-driven and hybrid approaches to product formulation can dramatically reduce development time and cost while improving knowledge retainment and expanding resource capabilities. The importance of spending time on these considerations should not be overlooked in favor of reaching fast preliminary modeling results (which may be misleading), nor should a dataset be quickly dismissed as insufficient without a meaningful and quantitative evaluation.

This presentation will discuss how to identify and overcome common pitfalls in formulation datasets, and will draw examples from various industries including polymers, specialty chemicals, and foods. ProSensus’ FormuSense software will be used to illustrate the typical steps required to optimally preprocess a raw formulation dataset for latent variable modeling and numerical optimization. Topics covered will include:

-structuring the raw data (such as identifying ingredient classes)
-detecting and resolving data anomalies (such as misspellings and missing ingredients)
-handling categorical variables (such as subject-matter expert knowledge)
-calculating ingredient class ratios and mixture properties to evaluate the impact of new ingredients
-meaningful statistics and visualizations to evaluate data suitability for modeling
-modeling approaches in the presence of missing data (such as raw material properties)

In the era of Industry 4.0, Big Data, and Digitalization, manufacturers are increasingly seeking opportunities to gain actionable insight from their process data. Some key product or process characteristics however, remain difficult to measure with conventional sensors or rely on expensive or time-consuming sampling and lab measurement.

Machine vision solutions are a mature technology that enable reliable, online measurement of visual and thermal product properties. These systems are capable of quantifying critical product characteristics such as texture or size distribution that would otherwise be difficult to measure.

This presentation will focus on three challenging use cases and will provide an overview of the use case, hardware details, algorithm details, and results for each application. The applications include quantifying:

-Texture of synthetic rubber crumb using color imaging
-Stirring intensity of hot metal using color videos
-Size distribution and moisture of iron ore using a polarization camera

Multivariate analysis (MVA) is a proven method for analyzing and interpreting large volumes of industrial data. MVA identifies the most significant correlations between input and quality variables, and the resulting prediction models can be leveraged to guide processes to desired outcomes. MVA models can be further exploited through the application of numerical optimization, where desired quality variables are achieved from an optimal combination of inputs, while maintaining the historical correlation structure of the data and relevant constraints.

MVA methods, such as projection onto latent structures (PLS), have been successfully used in optimization formulations owing to the common use of derivative-based algorithms. But with the rising popularity of black-box methods for prediction purposes (in particular, neural networks (NN)), it is now of interest to apply numerical optimization to these methods.The black-box nature of NN models raises some concerns surrounding the ability to find a feasible solution path and final solution. As such, the focus of this preliminary research is to assess the suitability of NN models for the application of numerical optimization, using an industrial polyolefin dataset. This dataset is used to build both a PLS model and a NN model to predict the same multiple quality variables. The PLS model is built directly from the raw input and quality data, whereas the neural network uses the PCA scores as a dimensionality-reduction preprocessing step prior to training. The two models are then formulated into optimization problems in order to solve for an optimal set of inputs based on a set of targeted quality variables.

The presentation will provide an overview of the concepts of MVA and how numerical optimization is applied to PLS models, followed by a discussion on how the PLS and NN models are trained with the dataset and their role in the optimization framework. Next, a comparison of the results of the model predictions and optimization cases will be presented. Finally, a discussion of the challenges encountered and next steps to further this research by moving forward with a dataset with more nonlinearities will be presented.