Time-series forecasting of demolding force in plastic injection molding using long short-term memory architecture

In this study, long short-term memory (LSTM) was used for real-time forecasting of the demolding force in an injection molding process. The melt temperature, mold temperature, filling time, and packing pressure were evaluated in three cases by using a quasi-experimental method. Three optimization algorithms, namely, stochastic gradient descent with momentum (SGDM), root mean squared propagation (RMSPROP), and adaptive moment estimation (ADAM), were used to train, test, and evaluate LSTM models based on time-series demolding force data. Model performance was evaluated using the root mean square error (RMSE) and coefficient of determination ( R 2 ). The results showed that all the optimization algorithms yielded satisfactory predictions, with the ADAM-LSTM delivering the best performance. Specifically, the ADAM-LSTM achieved a minimum RMSE of 3.55 N and a maximum R 2 of 99.45% under one specific condition, indicating excellent agreement between the predicted and actual demolding force profiles. These results confirm that the LSTM architecture, particularly when optimized with ADAM, effectively captures the nonlinear and temporal dynamics of demolding force in injection molding. In addition, the impact assessment of the time-step prediction validated the model’s ability to predict force progression over successive molding cycles. Overall, the proposed LSTM-based model provides a reliable tool for forecasting and monitoring the demolding phase in injection molding, enabling real-time process control and experimental validation.