A hybrid intelligence framework with uncertainty quantification for reliable surface roughness prediction in micro-milling

Abstract

In high-risk ultra-precision manufacturing, such as micro-milling, the application of artificial intelligence for predictive modeling and process optimization is critically hampered by a credibility gap stemming from the “black-box” nature of predictive models and the data-scarce reality of the domain. To bridge this gap, this study proposes a novel physics-constrained hybrid intelligence framework uniquely tailored to the physical realities of micro-milling. It is designed to deliver not just accurate, but fundamentally credible predictions from small sample data. The framework is built on a synergistic, three-tiered approach that explicitly mirrors the surface formation mechanisms: first, an explainable boosting machine model establishes an interpretable baseline with domain knowledge as soft constraints, which can capture macroscopic deterministic trends (e.g., kinematic geometry); second, a stacking ensemble method systematically corrects for complex, stochastic micro-dynamic residuals (e.g., size effect) to enhance accuracy; finally, the integration of adaptive conformal prediction—applied for the first time in this domain—augments point predictions with statistically rigorous uncertainty intervals, thereby transforming the model from a simple predictor into a trustworthy decision-support tool. Evaluated on experimental micro-milling data, the framework demonstrates its core value by achieving a 95.31% prediction interval coverage probability (PICP) at a 95% confidence level. This credibility is complemented by state-of-the-art accuracy, evidenced by a 19.9% reduction in mean absolute percentage error over the strongest baseline. By delivering accurate and uncertainty-quantified predictions, this framework presents a robust surface roughness predictive paradigm, providing manufacturers with reliable process-level decision support in data-scarce ultra-precision manufacturing environments.