Abstract
Objectives: This study aimed to develop and validate a machine learning-based predictive model for forecasting wear loss in additively manufactured (AM) dental resin materials using Long Short-Term Memory (LSTM) recurrent neural networks.
Materials and methods: Wear data were collected from three wear testing methods: Ball-on-Disc (BoD), Block-on-Ring (BoR), and Reciprocation (Recip), using three different AM resin materials under varying loads (49 N, 70 N) and surface treatments (polished, glazed). The LSTM model was trained on standardized time-series wear data using both Leave-One-Material-Out (LOMO) and Leave-One-Group-Out (LOGO) cross-validation strategies. Prediction windows were assessed at 10 %, 20 %, and 30 % of total wear sequences, simulating early-stage prediction of long-term wear progression. Model performance was evaluated using RMSE (Root-Mean-Square Error), MSE (Mean-Square Error), and MAE (Mean-Average Error).
Results: The autoregressive LSTM forecasting approach accurately predicted wear progression across all testing methods, with prediction accuracies ranging between 82 % and 97 % depending on method and dataset, the models explaining approximately 82-97 % of the wear variability (depending on method and dataset). Predictions initiated at 10 % showed high cross-validation accuracy, while test set generalization improved when prediction started at 30 %. Optimal model performance was achieved using a 50-point input window and step size. The model demonstrated robustness in handling the inherent variability of experimental wear data across multiple AM materials and test conditions.
Significance: This study demonstrates the feasibility of applying LSTM models for early and accurate prediction of wear progression in AM dental materials, offering potential for reducing physical testing duration and enhancing data-driven material evaluation frameworks for clinical applications.