How AI can better serve the chemical process industry
Years of production have accumulated rich data assets throughout the chemical process industry (CPI).
Years of production have accumulated rich data assets throughout the chemical process industry (CPI). In the age of artificial intelligence (AI), there is a strong momentum to use AI in process modeling, optimization, advanced control, debottlenecking, troubleshooting, etc., in case a first-principles-based approach cannot solve the problem effectively.
However, it may be very difficult to explain or interpret the resulting black-box AI models, and this hinders the trust and adoption of AI approaches. In this article, a brief introduction is provided of a novel method of “trustworthy AI (TAI)” for the CPI and its successful application in two chemical processes for improving operational excellence.
Process development is a daunting task for the CPI, which faces constraints on budget, time and manpower, and a limited number of pilot tests. Even though design engineers can use process simulation packages to run different scenarios or build in-house process models, the eventual process still could be suboptimal, and some important variables may still not be monitored and controlled. Additionally, some practical issues may not be exposed until the production stage, market conditions may change, and the original design purpose may not fit anymore. These factors indicate that room exists for further improvement in CPI performance metrics.
Most CPI plants accumulate years of data and records from the distributed control system (DCS), lab tests, enterprise information systems, etc. If fully utilized, tremendous value can be extracted from these gold mines. There are hundreds, even thousands, of variables in a process, and most of the relationships are non-linear. However, human cognition has a severely limited capacity: adult humans can retain only about four items “in mind.”1
Engineers and decision-makers always appreciate causal relationships and feel confident when they can see the direction guided by clear causal relationships. Based on sound principles in mass and energy balance, thermodynamics, reaction kinetics and mass transfer, first-principles models gain their trust easily. Simple linear regression models, like Eq. 1, can also be understood or explained easily.
y = a1 × x1 + a2 × x2 + ... + ai × xi + ... + an × xn (1)
Models and datasets
Based on transparency, interpretability and explainability, mathematical models are classified as “white-box” (or glass-box) models vs. black-box models.2 For example, linear regression models, linear logistic models (for classification) and simple decision trees are white-box models. Artificial neural networks (ANN, including deep learning), random forests (many decision trees) and support vector machines (SVM) are black-box models.
AI methods applied on large datasets are now influencing business decisions in various industries. However, the acceptance and application of AI methods are hindered since most of them are black-box models. It is often difficult to interpret these black-box models’ behavior and explain the results to users. As models become more complex, the task of producing an interpretable/explainable version of the model becomes more difficult.
For example, how can you tell someone, “You have a higher chance of developing a heart attack in the next two years,” or “We cannot approve your loan application,” just because the black-box model predicted it? In the CPI, how can you convince the plant to use a control system that changes the manipulated variables automatically just because the black-box model indicated that it should be so?
Raw datasets are often large and complex, with many anomalies and noises. Furthermore, in the CPI, due to the limited understanding of the intertwined reaction/mass transfer/energy transfer phenomenon, some key variables may not even be shown on the list of data tags or DCS dashboard, making it challenging to understand, evaluate and be confident about results of AI algorithms or models. Therefore, it is critical to develop new AI algorithms that can pinpoint the hidden key variables, clearly explain why and how AI algorithms perform and convince the domain experts in the CPI.
Explainable AI
In the past a few years, in the domain of computer science and data science, two types of explainable AI (XAI) methods were developed: post-hoc and ante-hoc. Ante-hoc techniques involve seeding explainability into a model right from the beginning. Post-hoc techniques continue with the black box phenomenon, where explainability is based on various test cases and “their results.”3,4 For highly complicated problems, post-hoc and ante-hoc methods may be combined to enhance the explainability of current AI algorithms. XAI has become increasingly popular in the health care industry and consumer business.
Even though XAI methods can explain the results of black-box models, they still cannot find and explain the importance of key variables (features) that are not yet shown in the list of data tags or on the DCS dashboard. So the AI models built—even with high accuracy—still can be flawed, and the results from XAI may just be a superficial explanation of a flawed model.
To overcome the limitations, the authors have developed a TAI method, as illustrated by FIG. 1. Rather than just using the variables documented in the existing data set and selecting key variables from them, TAI cleans the original data set first, then creates new variables that are undocumented and produces the associated data for these new variables based on the existing dataset. The new variables can be created from combination techniques, or from scientifically proven grouped numbers. For example, the Reynolds number (Re) is an important grouped number in fluid mechanics.
FIG. 1. Trustworthy AI for the chemical process industry.
Then, data analysis and model prediction algorithms will be applied to the expanded dataset, and the results will be explained accordingly. Decisions can be made after a convincing explanation is given.
CASE STUDIES
Plant DCS historian data was extracted, and the authors applied the TAI method to an ethylene oxide process and an ethanolamines process, respectively.
Case A: Ethylene oxide production
Ethylene is oxidized with oxygen to form ethylene oxide over silver alumina catalysts in parallel plug flow reactors (PFR) operated at 220°C–300°C (428°F–572°F) and 10 bar–30 bar pressure. Inhibitors are ejected to the reactors to prevent the generation of byproducts, mainly carbon dioxide (CO2) and water (H2O). The major reactions are shown in FIG. 2.
FIG. 2. Major reactions in ethylene oxide production.
The schematics of a typical ethylene oxide process5 are shown in FIG. 3.
FIG. 3. Schematics of a typical ethylene oxide process. Sections: 0) Make-up section. 1) Reaction section (multiple parallel reactors). 2) EO absorption section. 3) CO2 absorption bypass, purge and recompression section. 4) CO2 absorption section.
The plant wants to improve the selectivity of ethylene to ethylene oxide, rather than burning ethylene to CO2 and H2O. The ethylene feedrate and catalyst will remain the same. What actions should be taken? Where should capital be spent? Conventional wisdom focuses on the reaction systems. For example, increasing inlet O2 concentration, increasing inlet ethylene concentration, or adjusting the performance of the inhibitor are the most important aspects for the reaction system.
However, utilizing the TAI method, it was determined that the most critical and safe method to increase selectivity lies in the absorption section. By improving the absorption system, the selectivity can be increased by 3.3%, which yields $4.3 MM of profit increase.
Case B: Ethanolamines production
In an ethanolamines process, monoethanolamine (MEA), diethanolamine (DEA) and triethylamine (TEA) were produced from the reaction of ethylene oxide with aqueous ammonia solutions via sequential plug flow reactors with intercooling.6 A representative schematic diagram of the process is shown in FIG. 4.
FIG. 4. Schematic diagram of an ethanolamines process.
The reactions are very complicated, as shown in Eqs. 2–4:
EO+ NH3 → MEA (2)
EO + MEA → DEA (3)
EO + DEA → TEA (4)
Current market conditions favor only DEA production, so the plant wants to maximize DEA production and minimize the production of MEA and TEA. Admittedly, this is a difficult task since DEA is generated from the reaction between EO and MEA, but excessive EO will react with MEA to generate TEA.
Using the TAI method, the most important factors contributing to the selectivity of DEA were identified, and an optimal production strategy was identified, reducing the production rate of undesirable byproducts DEA and TEA by 66%.
Takeaway
TAI can augment traditional first-principles models and black-box models by providing the CPI with the following benefits:
- Discovery of possible new insights, new knowledge and ideas
- High fidelity
- Easier evaluation of AI results
- Increased confidence in adopting AI for application
- Easier maintenance of the AI algorithm or model.
These benefits will enable better process optimization, control, troubleshooting, debottlenecking, etc., which will help improve the performance of the CPI. HP
LITERATURE CITED
- Buschman, T. J., M. Siegel, J. E. Roy and E. K. Miller, “Neural substrates of cognitive capacity limitations,” Proceedings of the National Academy of the U.S.A., 2011, online: https://doi.org/10.1073/pnas.1104666108.
- Nori, H., S. Jenkins, P. Koch and R. Caruana, “InterpretML: A unified framework for machine learning interpretability,” Cornell University, 2019.
- Barredo Arrieta, A., N. Díaz-Rodríguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garcia, S. Gil-Lopez, D. Molina, R. Benjamins, et al., “Explainable artificial intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI,” Cornell University, 2019.
- Lundberg, S. M., P. G. Allen and S.-I. Lee, “A unified approach to interpreting model predictions,” University of Washington, 2017.
- Peschel, A., A. Jrke, K. Sundmacher and H. Freund, “Optimal reaction concept and plant wide optimization of the ethylene oxide process,” Chemical Engineering Journal, 2012.
- Algubury, H., A. Aljeboree, F. Karam and A. F. Alkaim, “Monoethanolamine: Production plant environmental pollution view project study kinetic and thermodynamic to release drugs atenolol and biporbiolol view project,” Research Journal of Pharmaceutical, Biological and Chemical Sciences, October 2015.
The Authors
Lou, H. H. - Lamar University, Beaumont, Texas
Helen H. Lou is a Professor in the Dan F. Smith Department of Chemical Engineering at Lamar University in Beaumont, Texas. Her areas of expertise include big data analytics, machine learning and process systems engineering. She has been applying these technologies for the sustainable development of the chemical processing industry. Dr. Lou has successfully implemented more than 50 projects sponsored by federal, state and private sectors, and has published extensively. She earned her MA degree in computer science and her PhD in chemical engineering, and has trained 16 master’s-level and 9 doctoral-level chemical engineers, as well as influenced many undergraduate chemical engineers. Due to her outstanding achievements, Dr. Lou was elected as a Fellow of the American Institute of Chemical Engineers (AIChE) in 2018.
Gai, H. - Lamar University, Beaumont, Texas
Huilong Gai is a research scientist at Lamar University. He received his PhD in chemical engineering from Lamar University in 2019. His work focuses on the modeling and control of petrochemical processes and combustion kinetic modeling and control of industrial flares. Dr. Gai specializes in various machine-learning algorithms and first-principles models.
From the Archive