Recurrent Neural Networks with Integrated Gradients Explanation for Predicting the Hysteresis Behavior of Shape...

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning Applications

Background:

• Shape memory alloys (SMAs) exhibit complex hysteresis behavior crucial for their applications.
• Predicting SMA hysteresis accurately is challenging due to nonlinearities and cycle-dependent effects.
• Recurrent neural networks (RNNs) offer a promising approach for modeling such complex material responses.

Purpose of the Study:

• To develop and evaluate recurrent neural network models for predicting the hysteresis behavior of SMAs.
• To compare the performance of SimpleRNN, LSTM, and GRU architectures in capturing SMA hysteresis.
• To assess the generalization and extrapolation capabilities of the trained models on independent data.

Main Methods:

• Dataset generation from 100-250 loading-unloading cycles across seven frequencies (0.1-10 Hz).
• Input features: applied stress (σ), cycle number (N), and loading-unloading stage (UpDown). Output: material strain (ε).
• Model training and validation using an 80/20 split (training/testing) with internal validation, employing LSTM, GRU, and SimpleRNN.

Main Results:

• All RNN architectures achieved excellent accuracy (R² > 0.999) on test sets.
• LSTM networks demonstrated the highest accuracy and most stable extrapolation across various frequencies.
• Integrated Gradient analysis revealed stress as the dominant factor, with cycle number indicating fatigue effects.

Conclusions:

• Recurrent neural networks, particularly LSTM, are effective for predicting SMA hysteresis.
• The models accurately capture complex behaviors and exhibit good generalization.
• Interpretability analysis confirms the physical relevance of the model's predictions, enhancing confidence in their application.