Deep learning approach for predicting the dynamic mechanical properties of epoxy-based shape memory polymers

Predicting the dynamic mechanical behaviour of shape memory polymer (SMP) is crucial for advanced applications, especially in space and aerospace industries, where temperature and vibration responses are critical. However, the existing literature on predicting dynamic mechanical analysis properties of epoxy-based SMP using deep learning is sparse. The challenges in performing time-consuming experiments, such as step-hold and multi-frequency tests, necessitate the exploration of alternative predictive methods. This study introduces a deep learning–based approach to predict the storage and loss modulus of epoxy-based SMP under varying temperature (25°C–150°C) and frequency (0.1–50 Hz) conditions. Experimental data, including dynamic strain, stress, time, temperature and frequency, were utilized as input for model training. Several deep learning models were evaluated, including core models – Convolutional Neural Networks (CNNs), Feedforward Neural Networks (FNNs) and Long Short-Term Memory Networks (LSTM), as well as hybrid models – Transformer_CNN (T_CNN), Transformer_FNN (T_FNN), Transformer_LSTM (T_LSTM), CNN_FNN, CNN_LSTM and FNN_LSTM. Among all models, the FNN demonstrated the best performance, particularly in predicting storage modulus within the mid-range of 400–500 MPa, while other models, including CNN and LSTM, failed to capture the trend, with predictions stagnating around 400 and 50 MPa, respectively. For loss modulus, the FNN and T_FNN models showed the highest accuracy across 0–60 MPa, whereas other architectures underpredicted values beyond 40 MPa. The FNN model demonstrated a root mean square error of 84.29 for storage modulus and 8.76 for loss modulus. Among the evaluated models, the FNN demonstrated the best overall performance, reducing the storage modulus prediction error by approximately 67.3% and 79.1% compared with the CNN and LSTM models, respectively. For loss modulus prediction, the T_FNN and FNN models reduced the prediction error by approximately 52.6% and 50.8%, respectively, compared with the CNN model. The results demonstrate the potential of deep learning models for predicting the dynamic mechanical behaviour of epoxy-based SMP.