Explaining Process Control Optimisation Recommendations via GradientSHAP and Implicit Differentiation

Explainability can bridge the trust gap in industrial automation: The paper addresses a real problem where engineers design optimization algorithms but operators hesitate to follow recommendations. By providing clear explanations of why an optimization system recommends specific actions, the approach helps operators understand and trust automated decisions in industrial settings.

Mathematical rigor enables practical speed improvements: The method uses the Implicit Function Theorem to calculate exact parameter sensitivities directly from optimization equations, achieving over 40× faster explanations compared to standard approaches while maintaining equivalent accuracy (>0.99 correlation with KernelSHAP). This speedup is critical for real-time industrial applications where operators need immediate answers.

Natural language explanations make technical outputs accessible: By combining mathematical attribution methods (GradientSHAP) with Large Language Model-generated narratives, the system translates complex optimization logic into language operators can understand. This bridges the gap between what algorithms compute and what humans can act upon.

The approach was validated on real industrial problems: Rather than testing only on toy examples, the researchers demonstrated their method on an actual High Pressure Grinding Roll (HPGR) control optimization problem with 22 input features and received feedback from domain experts, suggesting practical applicability beyond academic settings.