Comparison of Deterministic and Probabilistic Machine Learning Algorithms for Precise Dimensional Control and Uncertainty Quantification in Additive Manufacturing

ABSTRACT We present an accurate estimation of the dimensions of additively manufactured components by adopting a probabilistic perspective. Our study utilizes a previously gathered experimental dataset, encompassing five crucial design features for 405 parts produced in nine production runs. These runs involved two machines, three polymer materials, and two‐part design configurations. To illustrate design information and manufacturing conditions, we employ data models that integrate both continuous and categorical factors. For predicting Difference from Target (DFT) values, we employ two machine learning approaches: deterministic models that offer point estimates and probabilistic models generating probability distributions. The deterministic models, trained on 80% of the data using Support Vector Regression (SVR), exhibit high accuracy, with the SVR model demonstrating precision close to the process repeatability. To address systematic deviations, we introduce probabilistic machine learning methodologies, namely Gaussian Process Regression (GPR) and Probabilistic Bayesian Neural Networks (BNNs). While the GPR model shows high accuracy in predicting feature geometry dimensions, the BNNs aim to capture both aleatoric and epistemic uncertainties. We explore two approaches within BNNs, with the second approach providing a more comprehensive understanding of uncertainties but showing lower accuracy in predicting feature geometry dimensions. Emphasizing the importance of quantifying epistemic uncertainty in machine learning models, we highlight its role in robust decision‐making, risk assessment, and model improvement. We discuss the trade‐offs between BNNs and GPR, considering factors such as interpretability and computational efficiency. The choice between these models depends on analytical needs, striking a balance between predictive accuracy, interpretability, and computational constraints. In summary, our analysis of an additive manufacturing dataset through the lens of a Probabilistic Bayesian Neural Network (BNN) and the simultaneous quantification of both epistemic and aleatoric uncertainties provides a robust foundation for advancing manufacturing design.