Predicting ultimate tensile and break strength of SLS PA 12 parts using machine learning on tensile load–displacement data

Abstract Machine learning techniques are increasingly employed to predict material properties and reduce mechanical testing costs. This study uses tensile load‒displacement data to predict the breaking force, maximum force, ultimate tensile strength (UTS), and breaking strength of nylon PA 12 specimens manufactured via selective laser sintering (SLS). To address variations in specimen ductility, linear and fourth-degree polynomial equalisations were applied to the data. The results indicate that polynomial equalisation generally achieves superior predictive performance across various machine learning models. While all the models exhibit a decline in performance metrics with a reduction in the observation matrix data, several demonstrate strong predictive capabilities even with reduced data sets. Specifically, when using 20% of the observation matrix data (corresponding to the elastic zone) are used, the kernel ridge, elastic net, and Lasso regression models achieve R 2 values of 0.987, 0.931, and 0.987, respectively, for predicting breaking strength. For maximum force prediction with 20% of the data, the kernel ridge, support vector machine, and Lasso regression models resulted in R 2 values of 0.975, 0.960, and 0.930, respectively. However, the prediction accuracy for UTS and ultimate strength at break is lower, with the Elastic Net model achieving R 2 values of 0.703 and 0.682, respectively, with 20% of the data. These findings suggest that while certain mechanical properties can be accurately predicted using data solely from the elastic zone, minimising specimen waste and testing time, the prediction of other properties requires data beyond the elastic limit. This study highlights the potential of machine learning to optimise and streamline the quality control process in additive manufacturing.