Abstract Accurate prediction of creep behavior in additively manufactured polymers is essential for ensuring long-term structural reliability. This study presents a hybrid physics-informed long short-term memory (PI-LSTM) framework for modeling the creep and recovery behavior of additively manufactured PLA. The dataset consists of experimental creep data collected under varying stress levels (5 to 25 MPa), infill densities (70 to 100%), and infill orientations (0°, 45°, 90°), along with corresponding viscoelastic material parameters. In the proposed approach, the Burgers viscoelastic model is first used to generate physics-based strain predictions, which are then incorporated as input features into an LSTM network together with process parameters, time, and material properties. Unlike conventional physics-informed neural networks (PINNs), which embed governing equations into the loss function, the present method adopts a sequential physics-informed strategy in which the Burgers-model predictions provide physically guided information to the LSTM framework. The model was trained using Bayesian hyperparameter optimization, together with dropout regularization, gradient clipping, and early stopping based on validation performance to ensure good generalization and reduce overfitting. The proposed PI-LSTM achieves a prediction accuracy of 98% (R2) for the full dataset and 96% for an independent test set, outperforming the standalone analytical model and the conventional machine-learning model by approximately 22 and 14% in prediction accuracy, respectively. These results demonstrate that the hybrid framework improves predictive accuracy while maintaining physical consistency, making it a practical tool for digital twin development in polymer engineering applications.
