Sequence-to-Sequence Modeling of Hysteresis in Magnetorheological Elastomers Using Bidirectional LSTM Networks
Year: 2026
Type: conference-paper
Abstract: The study developed a memory-aware and quantitaive framework to predict the hysteretic behavior of Magnetorheological Elastomers (MREs) through Bidirectional Long Short Term Memory (BiLSTM) artificial neural networks. It is framed that the force- displacement dynamic relationship could be viewed as a sequence-to-sequence learning problem; therefore, the BiLSTM network generates force responses by taking displacement, velocity, frequency, and magnetic currents as inputs. In addition, the model has been verified through twelve experimental test conditions, and it accurately captured the non-linear viscoelastic effects of the MREs. Moreover, the predicted hysteretic loops exhibited closed loops and aligned trajectories when subjected to cyclic loading. Quantitatively, the stiffness error was less than five percent in over seventy percent of the test cases, and the minimum and maximum errors were 0.99 percent and eighteen point two six percent, respectively. Additionally, the BiLSTM architecture provided better energy loss representations and slightly better loop fidelity than traditional LSTM architectures. As such, this method will provide physically consistent predictions of MRE behaviors, and thus make it an appropriate candidate for smart damping and adaptive isolation applications.
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7201840955