Explainable AI-Based Intelligent Decision Support System for Sustainable Selection of Non-Traditional Machining Processes

Engineers choosing a non-traditional machining (NTM) process for a hard, brittle, or geometrically complex component typically rely on comparison tables, personal experience, and repeated trial runs. This is a reasonable approach given how differently processes such as electrical discharge machining, electrochemical machining, laser beam machining, and ultrasonic machining behave with respect to compatible materials, achievable accuracy and surface finish, and resource requirements — no single process dominates across all of these dimensions. Machine learning has recently been applied within individual NTM processes, mainly to predict material removal rate or surface roughness once a process has already been chosen, and a separate line of work has explored explainable AI (XAI) to make black-box manufacturing models more transparent. The earlier and arguably harder step — recommending which NTM process to use in the first place, explaining that recommendation in terms an engineer can act on, and weighing sustainability alongside the usual technical and economic criteria — has received comparatively little attention. This paper proposes a conceptual and partially computational decision support framework addressing all three elements together. Machining requirements — material, geometry, accuracy, surface finish, productivity, energy and environmental considerations, and cost — are used as input features for a tree-based classification model, with SHAP (SHapley Additive exPlanations) added to indicate which factors drive each recommendation, both across the model as a whole and for a specific job. A weighted sustainability score is proposed alongside the technical recommendation so that energy use, material consumption, and waste are treated as primary selection criteria rather than secondary concerns. The paper describes the proposed architecture, the data such a system would require, an evaluation approach suited to both predictive accuracy and interpretability, and the practical and data-related limitations any such system would face. No experiments have been conducted for this paper; its contribution is the framework itself, together with a literature-based argument for why it is needed.