Hierarchical Causal Abduction: A Foundation Framework for Explainable Model Predictive Control

Researchers present Hierarchical Causal Abduction (HCA), a framework that makes Model Predictive Control decisions interpretable by combining physics-informed reasoning, optimization evidence, and causal discovery. The method achieves 53% higher explanation accuracy than existing approaches across industrial control applications, addressing a critical barrier to deploying AI in safety-critical infrastructure.

Model Predictive Control has become essential for managing safety-critical systems like power grids, chemical plants, and building automation, yet the nonlinear optimization underlying these decisions remains a black box to human operators. This opacity creates significant deployment friction—operators cannot verify why a control action is safe or optimal, limiting adoption in regulated industries where auditability is mandatory. HCA addresses this fundamental trust problem by fusing three complementary evidence sources: domain knowledge graphs encode physical relationships, KKT multipliers reveal which constraints and objectives drove each decision, and causal discovery algorithms trace temporal dependencies. The framework's 53% accuracy improvement over LIME across three industrial domains demonstrates meaningful progress toward truly interpretable control systems. Critically, HCA operates without per-domain retraining, suggesting the approach captures generalizable principles about control decision-making rather than pattern-matching. The authors' validation with domain experts—not just algorithmic metrics—strengthens confidence that explanations match operator mental models. Ablation studies showing 32-37% accuracy drops when any component is removed indicate the framework's architecture is fundamentally sound; no single technique dominates. The extension beyond MPC to learning-based control hints at broader applicability in AI-driven automation. For industries managing critical infrastructure, this work reduces a major implementation barrier. However, success depends on whether operators truly find HCA's explanations useful in practice and whether 0.88 accuracy with calibration suffices for high-stakes decisions. The methodology's generalization to other prediction systems opens pathways for similar interpretability advances across autonomous decision-making.

• →HCA combines physics knowledge graphs, KKT optimization evidence, and causal discovery to explain nonlinear MPC decisions with 53% higher accuracy than LIME.
• →The framework requires no per-domain tuning for cross-domain accuracy, suggesting generalizable principles about control decision interpretability.
• →Expert validation across greenhouse, HVAC, and chemical engineering applications confirms explanations align with operator mental models.
• →All three evidence sources prove essential, with significant accuracy degradation when any component is removed.
• →The methodology extends beyond MPC to learning-based control and trajectory planning systems.