Published in the journal Advances in Bridge Engineering, the study introduces the Bouc-Wen Model-Magnetorheological Seismic Isolation Bearings-LSTM Analysis Model, abbreviated as BWMLA. This framework integrates a physical mechanical model with a data-driven neural network to address longstanding challenges in modeling the complex nonlinear responses of bridge components under seismic excitation. The primary objective was to resolve the limitations of traditional constitutive models, which often struggle to simultaneously capture hysteresis nonlinearity and viscoelastic frequency-dependent characteristics in magnetorheological materials.
The core innovation of the BWMLA framework lies in its construction of a parallel physical architecture that combines the Bouc-Wen model with the Kelvin model. The Bouc-Wen component is designed to characterize the nonlinear hysteresis behavior of the bearings, dynamically adjusting stiffness and damping based on magnetic field parameters. Meanwhile, the Kelvin model addresses the frequency-dependent viscoelastic properties that the Bouc-Wen model cannot independently describe. By connecting these two models in parallel, the researchers created a unified mechanical response equation that accounts for both the memory effects of hysteresis and the rate-dependent energy dissipation of the material. This structural integration allows for a more comprehensive representation of the bearing’s behavior under varying seismic conditions.
To enhance the predictive capabilities of this physical model, the study incorporates a Long Short-Term Memory, or LSTM, neural network. The LSTM module is tasked with capturing the complex temporal correlations between seismic motion inputs and the resulting bridge responses. Traditional numerical simulation methods often suffer from low computational efficiency when processing the nonlinear temporal responses of complex seismic motions. In contrast, the LSTM network utilizes gating mechanisms to preserve long-term temporal features and suppress gradient vanishing, enabling rapid prediction of bridge displacement and acceleration. This data-driven component maps the mechanical responses of the bearings to the overall dynamic behavior of the bridge, facilitating real-time analysis of seismic performance.
A critical aspect of the BWMLA framework is its phased collaborative optimization strategy, which synchronously tunes the parameters of both the physical model and the neural network. The optimization process begins with a genetic algorithm to adjust the Bouc-Wen and Kelvin mechanical parameters, minimizing the error in the physical model’s force predictions. Subsequently, Bayesian optimization is employed to fine-tune the LSTM hyperparameters, such as the learning rate and the number of hidden layer nodes. These two stages are executed alternately until the total loss function converges. This approach prevents the optimization process from falling into local suboptimal solutions, a common issue when coupling physical parameters with neural network hyperparameters. The adaptive weighting scheme ensures that the physical precision and predictive accuracy are balanced throughout the training process.
The study validated the BWMLA framework through both simulation testing and actual performance testing on a real-world structure. For the simulation phase, the researchers utilized the PEER NGA-West2 bridge dataset, which contains over 10,000 paired records of seismic structural responses across various bridge types. The simulation results demonstrated that the BWMLA model achieved a root mean square error of 9.2 kN for hysteretic force modeling, significantly outperforming traditional models such as the linear Kelvin-Voigt, Bingham, and Bilinear models. The frequency-domain energy storage modulus error was reduced to 5.1%, indicating a substantial improvement in the accuracy of predicting the dynamic stiffness of the bearings under rich frequency components.
The actual performance testing was conducted on a suspension bridge with a main span of 385 meters. This bridge features a combination of high-strength steel and concrete, with a tower height of 160 meters and a total mass of approximately 42,000 tons. The structure is equipped with a health monitoring system that provides comprehensive seismic response data, including triaxial acceleration, dynamic displacement time histories, and bearing reaction data. The researchers used 50 sets of seismic event data continuously collected by this monitoring system as an external verification set. This data remained completely separate from the model training process, ensuring an objective assessment of the BWMLA model’s generalization capabilities in a complex engineering environment.
In the real-bridge validation, the BWMLA model demonstrated superior predictive accuracy compared to benchmark models. The mean deformation displacement error was 19.7 millimeters, and the mean acceleration error was 0.30 meters per second squared. These values were lower than those produced by the Bilinear and Bingham models, which recorded higher errors in both displacement and acceleration predictions. The coefficient of determination for the response spectrum reached 0.91, indicating that the model can reproduce 91% of the bridge response spectrum energy. This high level of spectral matching is crucial for identifying dominant frequency components within seismic ground motion that may cause severe structural damage.
Computational efficiency was also a key focus of the study, as real-time seismic monitoring requires rapid data processing. The BWMLA model achieved an average computation time of 1.95 seconds during the actual bridge testing, which was lower than the 2.99 seconds recorded by the Bilinear model. The computational complexity was measured at 65.23 MFLOPs, and the model utilized 17.66% of the available memory. These metrics demonstrate that the phased collaborative optimization strategy effectively reduces iteration redundancy by decoupling parameter calibration from network training. The lightweight nature of the LSTM module, combined with the state sharing mechanism, allows the model to adapt to edge deployment scenarios, enabling the structural health monitoring system to deliver rapid post-earthquake safety assessments.
Sensitivity analyses were conducted to evaluate the robustness of the model parameters and hyperparameters. The researchers tested perturbations of up to 20% in the Bouc-Wen, Kelvin, and magnetic field coupling parameters. The results showed that the combined disturbance of the Bouc-Wen parameters significantly increased the root mean square error, highlighting the strong sensitivity of the model to hysteresis energy dissipation characteristics. However, the phased optimization strategy reduced the sensitivity to combined disturbances by 38%, supporting the robustness of the framework in complex scenarios. The analysis also revealed that the genetic algorithm convergence threshold had the most significant impact on overall accuracy, necessitating a careful balance between precision and computational efficiency.
Despite its promising results, the study acknowledges several limitations. The model validation was primarily based on data from a single-span suspension bridge, which restricts the direct application of the findings to non-standard structures such as cable-stayed or arch bridges. Additionally, the dynamic coupling mechanism of the parallel models in the high-frequency range remains imperfect, which may affect the prediction accuracy for dynamic loads rich in high-frequency components, such as those from near-fault earthquakes. The parameter coordination lag in the phased optimization strategy also constrains the model’s potential for real-time control applications. Future research will focus on conducting field validation across diverse bridge types and developing adaptive frequency-band decoupling algorithms to enhance the precision of the framework in predicting strong ground motions.
Subject of Research: Hybrid modeling of magnetorheological isolation bearings for bridge seismic analysis
Article Title: Magneto-rheological isolation bearings vibration damping performance in an actual bridge based on the Bouc-Wen-Kelvin parallel model and second-order optimized LSTM: a case study of a 385m suspension bridge
Article References: Cui, H. (2026). Magneto-rheological isolation bearings vibration damping performance in an actual bridge based on the Bouc-Wen-Kelvin parallel model and second-order optimized LSTM: a case study of a 385m suspension bridge. Advances in Bridge Engineering, 7(1), Article 15. https://doi.org/10.1186/s43251-026-00205-8
Image Credits: AI Generated
DOI: 10.1186/s43251-026-00205-8
Keywords: magnetorheological bearings, seismic isolation, LSTM network, Bouc-Wen model, bridge engineering, vibration damping, structural health monitoring, hybrid modeling, Magneto-rheological, isolation, bearings, vibration
News Source: Denise Maddox. (November 10, 2026). Hybrid Model Improves Seismic Damping Predictions on Suspension Bridge. Scienmag.



