Performance Improvement of Seismic Response Prediction Using the LSTM-PINN Hybrid Method.

Area of Science:

• Structural Engineering
• Artificial Intelligence
• Seismic Analysis

Background:

• Accurate prediction of structural responses to seismic loading is crucial for safety.
• Deep learning models like Physics-Informed Neural Networks (PINNs) and Long Short-Term Memory (LSTM) networks show promise but have limitations.
• PINNs lack long-term temporal dependency capture in nonlinear systems, while LSTMs lack physical interpretability.

Purpose of the Study:

• To develop a hybrid LSTM-PINN model that integrates the strengths of both LSTMs and PINNs.
• To enhance the prediction of dynamic structural behavior under seismic loading.
• To improve physical consistency and temporal dependency capture in structural response modeling.

Main Methods:

• A hybrid model combining LSTM for temporal learning and PINN for physics-based constraints was developed.
• The model was evaluated on single-degree-of-freedom (SDOF) and multi-degree-of-freedom (MDOF) systems subjected to the El-Centro ground motion.
• Performance was assessed using mean error and mean squared error (MSE) for displacement, velocity, and acceleration, compared to PINN-only models.

Main Results:

• The hybrid LSTM-PINN model achieved more stable and precise predictions across the entire time domain.
• It demonstrated superior performance over baseline PINNs, achieving up to 50% lower MSE with fewer training epochs (10,000 vs. 50,000).
• The model showed improved generalization capabilities through temporal sequence learning.

Conclusions:

• The hybrid LSTM-PINN model effectively combines temporal learning and physical consistency for seismic structural response prediction.
• This physics-guided time-series AI approach offers significant advantages in accuracy and efficiency.
• The findings support the potential for real-time response estimation, structural health monitoring, and seismic performance evaluation.