<p>Optimizing the Hall–Heroult process requires balancing the partially competing objectives of energy minimization and process stability. Traditional control strategies often treat these two goals as strongly coupled, yet their degree of dependence has not been quantitatively clarified. This study proposes a data-driven analytical framework to distinguish the factors associated with energy efficiency from those associated with process stability and to derive interpretable control recommendations. By integrating unsupervised clustering, cluster-stratified causal inference, Bayesian structure learning, and scenario-based counterfactual simulation on observational industrial data, we find that energy-based and stability-based operating modes are nearly independent from the perspective of cluster consistency (Adjusted Rand Index <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\approx\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>≈</mo> </math></EquationSource> </InlineEquation> 0). Causal effect estimates are interpreted under stated identifying assumptions within each operational cluster and the results suggest that energy consumption is more closely associated with aluminum output volatility, whereas process stability is more strongly related to fluoride, temperature and alumina fluctuations. Based on these findings, a policy tree is further used to compare candidate interventions under different optimization objectives. The proposed approach provides an interpretable basis for separating energy-oriented and stability-oriented operational decisions in aluminum electrolysis cells.</p>