Causal Discovery for Topology Reconstruction in Industrial Chemical Processes

This paper explores the application of causal discovery frameworks to infer the topology of industrial chemical processes, which is crucial for operational decision-making and system understanding. While traditional data-driven methods entail process interventions, causal discovery offers a noninvasive approach. Challenges such as temporal aggregation, subsampling, and unobserved confounders, which can lead to false predictions, are emphasized in the paper. Through simulation case studies, the performance of various causal discovery methods under different observation scenarios is evaluated. Our findings underscore the importance of simultaneously considering instantaneous and lagged causal relations, highlight the suitability of structural equation modeling for temporally aggregated processes, and caution against misinterpretation of subsampled data. Additionally, we demonstrate the utility of the Wiener separation in identifying unobserved confounders, which is essential for navigating the complexity of industrial processes.