Artificial Intelligence and Process Systems Engineering

Abstract

Abstract Process systems engineering is a thriving field within chemical engineering. PSE deals with several design and operational tasks that allow process systems to work efficiently and safely. There is a large intersection between PSE tools and Artificial Intelligence (AI) algorithms, recognized for decades now. With unprecedented availability of various forms of data and significant improvement in computational prowess, AI techniques have started to address large and meaningful engineering problems. In this talk, we will explore the relevance and importance of AI techniques in the next generation process systems engineering applications. Various aspects of PSE and the impact of AI cross-cutting these aspects will be described as outlined below. The focus of this talk will be on the most recent developments and industrial applications that the author has been involved in. PSE as an area has implications in process modelling, process design, process optimization and process operations. Computer-aided tools are at the centre of all modelling activities. With the advent of AI, automated model building tools are being increasingly researched. Assembling first principles models in a purely data-driven manner is a promising area. Of course, process design is a key aspect of PSE. Design is an inverse problem, where a set of requirements are provided and designs that can satisfy the requirements are desired. As a result, any data-driven modelling tool can also be used in design if there are many exemplar designs that are available for training. As the result, the strength of AI in modelling can be leveraged for this inverse modelling problem. Natural evolution inspired techniques such as genetic algorithms also continue to play an important part in addressing complicated inverse design problems. Recently, reinforcement learning has also been used in solving design problems. The use of AI techniques in optimization is another exciting area of research. Many core AI algorithms themselves use optimization techniques in their development; use of learning approaches in optimization is an interesting synergy between the two fields. Convex representations using neural networks that allow convex optimization approaches to be used in optimization is an emerging area of research. Other convenient representations from an optimization viewpoint are likely to pursued. An example of such a representation is the difference of convex representation. The biggest impact of AI in PSE is in the area of process operations. With the ability of systems to collect data at an unprecedented level and the possibility of collecting variegated datatypes, AI algorithms can now be comprehensively explored for various process operations tasks. In process monitoring and operator training, natural language processing ideas have a large role to play. Further, data from different types of sensors such as vision, noise and so on, over and above the standard sensor data, is likely to revolutionize the way process monitoring and fault detection and diagnosis tasks are performed. This is particularly powerful when data from different plants are centralized allowing for the possibility of transfer learning to occur. Standard data rectification and gross error detection techniques that used to rely on process models are now being addressed by purely data driven approaches. This brings in several important questions that need to be satisfactorily addressed by the machine learning techniques. Interestingly, sensor placement for data reconciliation, fault detection and diagnosis algorithms that used to rely on process models are also being reimagined as data driven problems. Work on the use of neural networks and knowledge-based systems in control has been around for more than three decades. However, with renewed interest in AI, these approaches are being explored again with better architectures and larger computational power. Reinforcement learning is a natural approach to address several learning-based control problems. There has been a flurry of activity in this area, and one would expect this area to progress quite rapidly. There are several challenges related to inclusion of constraints, robustness and so on that need to be addressed comprehensively. Looking forward, two important streams of work can be identified. One of those is the hybridization of existing knowledge with the data driven AI systems. This will be a very profitable area of research and will bring in systems that are explainable, robust and more deployable in engineering problems. Another avenue that will assume significance is moving towards purely unsupervised learning. Many successful applications use supervised and/or semi-supervised learning approaches. However, in the future, several concepts for unsupervised learning will be explored. This, we believe, will lead to truly intelligent process systems that are safe, efficient and robust to inherent variations that cannot be controlled.