(632e) Causal Discovery in Chemical Processes: Dealing with Cycles and Latent Confounders

Chemical process systems involve intricate, large-scale interactions among various components, consisting of complex dynamics and feedback loops. Furthermore, process integration, intensification, and technological advancement in industrial equipment increase the scale and complexity of chemical process systems. First-principles models are often leveraged to describe these interactions. However, deriving, validating, and applying such models for control and optimization becomes increasingly difficult as processes become more complex. In addition, the aim toward increased automation in industrial operations motivates pursuing methods that automatically identify such complex interactions.

Causal discovery provides an alternative framework to traditional system identification for identifying complex interconnections between chemical process variables. It does not require process intervention or deliberate perturbation of the system from normal operating conditions [1]. In causal discovery, chemical processes are represented as graphs with nodes (or vertices) corresponding to process variables and edges corresponding to dynamic interactions [2,3]. Using this graphical representation, we can identify the underlying causal mechanism of the process and analyze how information propagates through the system. Previous works have applied causal discovery frameworks to chemical process systems in the context of root-cause diagnosis, fault detection, and topological reconstruction of chemical process systems [4]. However, the majority of these causal discovery techniques are limited to detecting acyclic causal relationships without unobserved common causes (or latent confounders), while feedback and cyclic interactions are ubiquitous in chemical processes [5,6,7,8]. Furthermore, complete observation of all possible process variables is often unattainable and expensive, leading to the problem of unmeasured common causes that can induce false causal predictions. The causal discovery of cyclic processes with latent confounders is further complicated by the presence of both lagged and very fast (essentially instantaneous) interactions between variables. Therefore, accounting for cycles and unmeasured common causes is critical in obtaining an accurate causal representation of the system.

In this talk, we will highlight the substantial rise in complexity encountered when transitioning from acyclic to cyclic causal discovery of chemical process systems with a mix of instantaneous and lagged causal interactions, particularly in the presence of latent confounders. We will elaborate on a graphical representation that is robust towards cycles and latent confounders known as partial ancestral graphs (PAGs) and provide insights that are critical to the correct causal interpretation of PAGs through realistic case studies [7,9].

We will first consider the causal discovery of a chemical process network consisting of reactors, a separation unit, and a recycle between the reaction and separation section. We will apply a broad class of causal discovery algorithms that are robust towards cycles and latent confounders to the simulated time series data of the process and analyze the accuracy of each algorithm. Using PAGs, we will illustrate how conditional independence information and independent component analysis can be leveraged to identify cyclic causal structures and unmeasured common causes and discuss the general challenges of cyclic causal discovery and effects of latent confounders by benchmarking the performance of these methods on this data set [7,9,10,11,12]. We will also demonstrate the application of the same set of causal discovery algorithms on a large-scale mining and pyrometallurgy operation process, the Arc-Loss dataset, to identify the cause for a sudden power loss in furnace plasma arc [13].

References

[1]. Ljung, Lennart. System identification. Birkhäuser Boston, 1998.

[2]. Pearl, Judea. Causality. Cambridge university press, 2009.

[3]. Bauer, Margret, et al. "Finding the direction of disturbance propagation in a chemical process using transfer entropy." IEEE transactions on control systems technology 15.1 (2006): 12-21.

[4]. Vuković, M. and Thalmann, S., 2022. Causal discovery in manufacturing: A structured literature review. Journal of Manufacturing and Materials Processing, 6(1), p.10.

[5]. Glymour, Clark, Kun Zhang, and Peter Spirtes. "Review of causal discovery methods based on graphical models." Frontiers in genetics 10 (2019): 524.

[6]. Vowels, M.J., Camgoz, N.C. and Bowden, R., 2022. D’ya like dags? a survey on structure learning and causal discovery. ACM Computing Surveys, 55(4), pp.1-36.

[7]. Zhang, J., 2008. On the completeness of orientation rules for causal discovery in the presence of latent confounders and selection bias. Artificial Intelligence, 172(16-17), pp.1873-1896.

[8]. Shimizu, S., Hoyer, P.O., Hyvärinen, A., Kerminen, A. and Jordan, M., 2006. A linear non-Gaussian acyclic model for causal discovery. Journal of Machine Learning Research, 7(10).

[9]. Richardson, T.S., 2013. A discovery algorithm for directed cyclic graphs. arXiv preprint arXiv:1302.3599.

[10]. Spirtes, Peter, et al. Causation, prediction, and search. MIT press, 2000.

[11]. Lacerda, G., Spirtes, P.L., Ramsey, J. and Hoyer, P.O., 2012. Discovering cyclic causal models by independent components analysis. arXiv preprint arXiv:1206.3273.

[12]. Materassi, Donatello, and Murti V. Salapaka. "Reconstruction of directed acyclic networks of dynamical systems." 2013 American Control Conference. IEEE, 2013.

[13]. Yousef, I., Rippon, L.D., Prévost, C., Shah, S.L. and Gopaluni, R.B., 2023. The arc loss challenge: A novel industrial benchmark for process analytics and machine learning. Journal of Process Control, 128, p.103023.