Enhanced Counterfactual Explanations for Optimizing Three-Dimensional Printing Parameters Using SHAP and Nearest-Neighbor Constraints With Physics-Based Validation

Abstract Overheating during metal 3D printing, particularly in laser-based powder bed fusion, can negatively impact the quality and reliability of printed parts. This study presents a novel framework for recommending printing parameters such as power, speed, and scanning sequence that reduce the risk of overheating. The method is adopted from the field of explainable artificial intelligence (XAI), specifically counterfactual explanation (CE), and is enhanced to support quality and feasibility in engineering applications. Our method, SHapley Additive exPlanations (SHAP) and neighbor-enhanced counterfactual explanations (SHANCE), integrates feature importance with domain-specific constraints to ensure interpretable and actionable recommendations. It identifies the most influential parameters driving overheating and proposes minimal, valid changes. The framework also includes a physics-based constraint on energy input to ensure parameters remain within viable operating ranges. In addition, a similarity-based filtering mechanism is incorporated to maintain realism. We validate SHANCE on a standard image classification dataset and a real-world 3D printing dataset. Results show that SHANCE generates diverse and realistic recommendations, with the least magnitude of change required compared to other existing CE methods, in classification tasks, and effectively mitigates overheating in laser powder bed fusion (L-PBF) by identifying optimal printing parameters. Compared to baseline optimization methods such as genetic algorithms and a neural network-based surrogate approach, SHANCE offers a balanced tradeoff between interpretability, feasibility, and performance. This makes it a robust tool for parameter optimization in metal additive manufacturing.