From feedforward to quantum: Exploring neural networks for predicting tensile strength in additively manufactured polylactic acid parts

Abstract

Accurate prediction of ultimate tensile strength (UTS) in additively manufactured parts is critical for optimizing process parameters and ensuring reliability. Although neural networks (NNs) have shown potential in modeling complex nonlinearities, most prior studies relied mainly on conventional feedforward neural networks (FFNNs), overlooking more advanced architectures. This work presents a comparative study of eight NN models, grouped into four categories: traditional (FFNN), memory-based (LSTM, GRU, Simple RNN), transformer-based (TabTr, FTTr, TabNet), and quantum-based (QNN). A dataset of 366 samples compiled from 22 published studies on Polylactic Acid (PLA) parts was used to evaluate these models. Additive manufacturing involves inherent yield constraints, such as reduced UTS with higher raster orientation or layer thickness, and the need to balance nozzle temperature and printing speed to avoid poor bonding or thermal degradation. By addressing these constraints through predictive modeling, the proposed approach provides a scientific contribution to experimental design and yield optimization, enabling reduced trial-and-error and material waste. Among the tested models, the Feature Token Transformer (FTTr) achieved the highest accuracy (R² = 90.08 %, RMSE = 3.041). To further interpret its predictions, Shapley Additive Explanations (SHAP) were employed, revealing raster orientation, infill density, and nozzle diameter as the most influential parameters. These findings confirm the robustness and interpretability of transformer-based networks for modeling UTS in additively manufactured PLA, underscoring the broader value of integrating modern architectures with explainable artificial intelligence to advance experimental design and optimization in additive manufacturing research. © 2025 The Authors