This study evaluates the effectiveness of three causal discovery algorithms in uncovering relationships within real-world manufacturing datasets, with a focus on supporting transparency and sustainability. Two constraint-based methods—Peter-Clark (PC) and Fast Causal Inference (FCI)—and one score-based method—Greedy Equivalence Search (GES)—are compared. Unlike traditional machine learning models that emphasize pattern recognition, causal models aim to uncover underlying cause-and-effect structures, enhancing interpretability and enabling more informed decision-making. The algorithms were applied to two high-dimensional datasets obtained from a German small-to-medium enterprise (SME) specializing in customized steel products. The datasets capture process duration and power consumption, two key indicators of production efficiency and energy use. Initial causal graphs were constructed with domain experts and used as references to evaluate algorithm outputs, which were further validated through bootstrap resampling. By identifying the root causes of energy consumption and production variability, this study supports data-driven interventions for process optimization and sustainable manufacturing. The results highlight the trade-offs between robustness and interpretability across methods, with PC producing the most stable and domain-aligned structures. Overall, this work demonstrates the practical value of causal discovery for advancing explainable and sustainable decision-making in industrial contexts.
Jacob et al. (2026) studied this question.