Comparative study of predictive models for the ultimate tensile strength of FDM-printed PLA parts based on literature-integrated data

Abstract

Fused deposition modeling is an important additive manufacturing technology, and accurate prediction of the mechanical properties of three-dimensional-printed parts is essential for engineering applications. However, existing studies are often limited by small sample sizes, insufficient comparison of predictive methods and a lack of external data validation. This study therefore aims to construct a cross-literature data set and compare different predictive approaches for the ultimate tensile strength (UTS) of polylactic acid printed parts. A total of 235 experimental data points from ten published studies were integrated to establish a data set for UTS prediction. Five input variables were considered: infill density, nozzle temperature, nozzle diameter, layer height and printing speed. Response surface methodology, multiple machine learning regression algorithms and an artificial neural network (ANN) were systematically compared under unified preprocessing conditions. In addition, ensemble models were constructed using histogram-based gradient boosting regressor (HGBR), gradient boosting machine (GBM), random forest and kernel ridge regression (KRR). External validation was performed using 36 independent data points. Among the single models, HGBR achieved the best performance with a test-set R² of 0.8952. The GBM-KRR-HGBR ensemble further improved accuracy, reaching a test-set R² of 0.9290. For this ensemble model, external validation showed that 61.11% of samples had prediction errors below 10%. Permutation importance analysis indicated that infill density was the most influential variable. The originality of this study lies in a reliability-oriented literature-data curation strategy, a unified comparison of statistical, machine learning, ANN and ensemble models under the same data framework, and a source-wise external error analysis for evaluating model applicability under heterogeneous literature-derived data conditions.