Manufacturing Data Science: From Explainable AI to Data-Driven Decision-Making

Abstract

Manufacturing data science leverages machine learning to support process monitoring, quality prediction, and operational optimization. However, predictive accuracy alone is insufficient for effective decision-making, particularly when model explanations lack fidelity and stability, uncertainty is inadequately characterized, or operational decisions must account for heterogeneous manufacturing conditions.

This talk presents our recent research on explainable artificial intelligence (XAI) and data-driven decision-making, with an emphasis on statistical machine learning and interpretable optimization. First, we investigate the limitations of Local Interpretable Model-agnostic Explanations (LIME) in local fidelity and explanation stability by incorporating multivariate adaptive regression splines and bootstrap aggregation, while employing Bayesian inference to quantify uncertainty in feature contributions. Next, we address manufacturing process parameter optimization through robust and interpretable policy learning. By integrating cross-fitting, doubly robust estimation, and policy forests, the proposed framework learns context-specific decision policies from observational manufacturing data. To enhance decision interpretability, local surrogate trees are employed to translate complex policy recommendations into transparent, instance-specific decision rules. Finally, we discuss future directions in trustworthy manufacturing data science, focusing on explainability, uncertainty quantification, and adaptive decision-making.

Please click here for participating the talk online.