Robust and Interpretable Policy Learning for Manufacturing Process Parameters

In manufacturing systems, process parameters (controllable factors) strongly influence quality outcomes, yet substantial instance-level heterogeneity—arising from environmental conditions, material variability, and machine states—often leads to context-dependent responses to parameter adjustments. Under such heterogeneity, traditional experiment-based approaches are costly and difficult to scale, whereas purely predictive machine learning models may produce unreliable decision policies when trained on observational data. We address this challenge by formulating process parameter optimization as a robust and interpretable policy learning problem inspired by causal inference principles. The proposed framework integrates preprocessing and action encoding, doubly robust policy value estimation to mitigate nuisance model misspecification, and policy optimization tailored to heterogeneous manufacturing contexts. To enhance interpretability, we introduce a policy forest for robust policy learning and apply LIMEtree-based distillation to approximate the learned policy with shallow instance-wise surrogate trees. This step produces transparent decision rules while maintaining fidelity to the