AIChE 2023 Spring Meeting

Cracking furnaces are a significant energy consumer and consequently there are many long-term sustainability projects focusing on physical design changes that will result in energy reduction. However, the goal of this work was to quickly minimize energy consumption in Braskem’s existing furnaces without incurring capex to make physical plant changes. Multivariate analysis can be used to analyze historical operating data to identify opportunities to reduce energy consumption while maintaining throughput and adhering to other manufacturing constraints. This project was achieved through a combination of offline modeling, real-time multivariate process monitoring, and real-time optimization. This presentation will provide an introduction to multivariate analysis, an overview of the Braskem’s approach using AspenTech’s ProMV suite, and discuss some results achieved.

For manufacturers, the continuous development of new products is necessary to remain competitive, whether it is to reduce cost, utilize alternative materials, or to meet new regulations. A design-of-experiments approach may not be the most effective tool to explore formulation options, especially for products that contain many ingredients and properties. Instead, to capitalize on data collected during product development, multivariate analysis (MVA) can be used to build a model to predict final product quality based on ingredient mixture properties, along with other fundamental knowledge.

In tandem with these statistical models, there are factors to consider that may not be captured in the model correlation structure, such as the importance of cost and quality, the introduction of new ingredients, and constraints on ingredient ratios, classes, and mixture properties. In order to explore solutions of this multifaceted problem, a projection onto latent structures (PLS) model is trained and embedded in a nonlinear constrained optimization problem.

This presentation will provide a walkthrough of the embedded optimization problem, how modeling decisions are considered in the formulation, and practical solution approaches.

Multivariate modeling (PLS) is a proven modeling method that is known to handle the highly correlated variables and missing data challenges often present in product development datasets. Additionally, datasets of various sizes can be effectively modeled and inverted using constrained numerical optimization. In this presentation, both PLS modeling, simulation, and numerical optimization will be applied to two datasets with the goal of developing new or reformulating existing products.

These two applications include:
-Development of new LDPE resins with targeted properties
-EPDM Rubber compound adhesion modeling and design of targeted experiments

In both datasets the following will be highlighted with the use of ProSensus’ FormuSense software:
-The appropriate modeling structure given the data available
-Modeling results and model limitations due to dataset limitations
-The design of new experiments or products through simulation or optimization

Gone are the days where a small group of domain experts plan single factor or simple mixture DOEs, and then patiently wait for these experiments to be scheduled into busy laboratories or pilot plants. R&D teams across many industries are embracing AI-powered tools to transform workflows and accelerate innovation.

This presentation will demonstrate the benefits of designing experiments with FormuSense. ProSensus developed FormuSense to be a streamlined implementation of our proven product development framework, which is centered around multivariate analysis.

A foods dataset (of baked goods) will examined in FormuSense to discuss the intuitive visualizations and features available that help formulators rapidly understand key correlations among and between ingredient properties, formulation ratios, process conditions, and final quality. The no-code platform will be used to suggest a small number of useful experiments to explore new or underrepresented regions of the design space.