Uncertainty Quantification in Machine Learning Models for Additive Manufacturing: A Bayesian Approach to Enhancing Model Robustness and Trustworthiness

Abstract

Machine learning approaches have become increasingly prevalent in the optimization and control of additive manufacturing processes over the past decade. Despite their widespread adoption, quantifying uncertainty in these models remains a significant challenge for ensuring reliable predictions in critical manufacturing applications. This paper presents a comprehensive Bayesian framework for quantifying uncertainty in machine learning models specifically tailored for additive manufacturing processes. We develop a hierarchical probabilistic approach that captures both aleatoric uncertainty arising from inherent process variability and epistemic uncertainty stemming from model limitations and data scarcity. Our methodology integrates Gaussian process regression with Markov Chain Monte Carlo methods to provide robust uncertainty estimates across diverse additive manufacturing scenarios. Experimental validation on laser powder bed fusion processes demonstrates that our approach reduces prediction error by 37\% compared to deterministic methods while providing well-calibrated uncertainty bounds. Furthermore, the proposed framework enables adaptive sampling strategies that optimize material property predictions with 42\% fewer experiments. This work establishes a foundation for uncertainty-aware decision-making in additive manufacturing, enhancing process reliability and accelerating qualification procedures for critical components.