Evolutionary computation at Dow Chemical

Abstract

Dow Chemical is the largest US chemical company by sales ($46 billion in 2005). It is one of the most diversified global company in the business with more than 3000 products in the market, mostly plastics, polyurethanes, and different specialty chemicals. There are two key directions in introducing new emerging technologies, such as Evolutionary Computation (EC), in the company. The first direction is continuous improvement of manufacturing and supply-chain efficiency, which becomes critical in the current economic environment of skyrocketing energy prices. The second direction is faster invention of an attractive pipeline of new products, which is critical for the future competitiveness of the company. The more specific issues of the chemical industry in these two directions can be defined as:• high-dimensionality of plant data (thousands of variables and control loops),• scarcity of data in new product development,• low speed of data analysis in High-Throughput Screening (which is at the basis of new product discoveries),• increased requirements for model robustness toward process changes due to the cyclical nature of the industry,• multiple optima,• key process knowledge is in process operators and poorly documented,• high uncertainty in new material market response,• supply chain not as well developed as process modelingThere are tremendous opportunities for EC to satisfy these needs and contribute to process improvement and new product discovery. However, we need to take into account that as a result of the intensive modeling efforts in the last 20 years many manufacturing processes in the most profitable plants are already supplied by different types of models (steadystate, dynamic, model predictive control, etc.). This creates a culture of modeling "fatigue" and resistance to introduction of new solutions, especially based on unknown technologies. This makes the efforts of applying EC in the chemical industry especially challenging since demonstration of significant competitive advantages relative to the alternative modeling and hardware solutions is required.