(60f) Explainable Artificial Intelligence (XAI)-Based Causality Analysis for Chemical Process

Machine learning (ML)-based surrogate models are simplified models that approximate nonlinear systems using input-output data[1]. They can capture the nonlinear relationship between the inputs and outputs of chemical processes and provide high predictive performance. Using surrogate models, researchers can assess the economic feasibility and environmental impact of different process alternatives[2] or identify data points that deviate from normal behavior[3,4] in process systems engineering (PSE). However, ML models are known as black-box models, which means that the user cannot easily know the reason for the model’s decision[5]. Although there is a growing demand of ML models for chemical processes with increased nonlinearity[6], the opacity of these models prevents them from being applied to industry.

To solve this problem, there has been an increasing number of studies to enhance the interpretability and trustworthiness of artificial intelligence (AI) models, which is called explainable artificial intelligence (XAI). In PSE, XAI can be applied to unsupervised fault diagnosis and detection[3,4,7], and process quality[8]. The limitation of these traditional methods is that they require the knowledge of engineers to provide clear feedback on the underlying cause. Therefore, AI-based process causality analysis is more applicable to understanding the cause and effect to optimize the process metrics globally and locally.

Here, we developed a methodology based on Shapely flow[9] that provides comprehensive information on the parameters that affect the technoeconomic analysis (TEA) and life-cycle assessment (LCA) of the chemical process. We can obtain the quantitative and qualitative feedback and visualize the relationships between process variables in a causal graph. This research contributes to a high-performance surrogate modelling and root cause identification to optimize CO2 emission and product yield at a reasonable cost.

[1] NENTWICH, Corina; ENGELL, Sebastian. Application of surrogate models for the optimization and design of chemical processes. In: 2016 International Joint Conference on Neural Networks (IJCNN). IEEE, 2016. p. 1291-1296.

[2] CHUNG, Wonsuk; LEE, Jay H. Input–output surrogate models for efficient economic evaluation of amine scrubbing CO2 capture processes. Industrial & Engineering Chemistry Research, 2020, 59.42: 18951-18964.

[3] JANG, Kyojin, et al. Explainable Artificial Intelligence for Fault Diagnosis of Industrial Processes. IEEE Transactions on Industrial Informatics, 2023.

[4] JANG, Kyojin, et al. Adversarial autoencoder based feature learning for fault detection in industrial processes. IEEE Transactions on Industrial Informatics, 2021, 18.2: 827-834.

[5] GUNNING, David; AHA, David. DARPA’s explainable artificial intelligence (XAI) program. AI magazine, 2019, 40.2: 44-58.

[6] DOBBELAERE, Maarten R., et al. Machine learning in chemical engineering: strengths, weaknesses, opportunities, and threats. Engineering, 2021, 7.9: 1201-1211.

[7] BRITO, Lucas C., et al. An explainable artificial intelligence approach for unsupervised fault detection and diagnosis in rotating machinery. Mechanical Systems and Signal Processing, 2022, 163: 108105.

[8] SENONER, Julian; NETLAND, Torbjørn; FEUERRIEGEL, Stefan. Using explainable artificial intelligence to improve process quality: Evidence from semiconductor manufacturing. Management Science, 2022, 68.8: 5704-5723.

[9] WANG, Jiaxuan; WIENS, Jenna; LUNDBERG, Scott. Shapley flow: A graph-based approach to interpreting model predictions. In: International Conference on Artificial Intelligence and Statistics. PMLR, 2021. p. 721-729.