(389a) A Review and Outlook on Machine Learning and Data Science in Chemical Engineering | AIChE

(389a) A Review and Outlook on Machine Learning and Data Science in Chemical Engineering


Qin, J. - Presenter, University of Southern California
Dong, Y., University of Southern California
With today’s development of the internet, internet of things, wireless sensors and communications, smart manufacturing, and high-throughput experiments, the amount of data collected has grown exponentially, analogous to the Moore’s law in the increase of solid-state transistor density. The data explosion has made all sectors including engineering, medicine, business, commerce, finance, and materials science to endorse the potential and power of big data. The relevance of big data to chemical engineering and science is evident from the recent technical contributions by chemical engineers and scientists.

Great successes in machine learning today have to do with natural language processing (NLP), multi-lingual translation, image recognition, etc. In addition, the demonstration of AlphaGo by Google's DeepMind in early 2016 was one of the most shocking events to the world. However, the game of Go has little in common with the degree of uncertainty and the complexity of real-world engineering problems. In this talk we discuss the characteristics of process systems engineering problems and provide a brief review of machine learning and data analytics in chemical engineering in the last three decades. We give a few attributes for machine learning techniques to be widely applicable in the process systems. Owing to these characteristics, process monitoring and inferential sensors are two of the most adopted data-driven applications in chemical engineering.

The talk will then provide an assessment of the resurgent interest in deep neural networks and deep reinforcement learning and contrast them against the technology two decades ago. We also give a few examples of high-profile applications that failed the test of time. We assert that most of the sustainable applications are domain-specific, in that they incorporate specific structures and knowledge in the models. We will briefly touch on the tremendous advancement in statistical machine learning, which are the core of data science and could propel the advances in other scientific and engineering disciplines, such as smart manufacturing, bioinformatics and materials science.