Rocking-response-prediction: A Seismic Response Prediction Model Based on Deep Neural Networks

Rocking-response-prediction is a data-driven method based on deep neural networks, used to predict the response of rocking rigid bodies under seismic action. It predicts whether a building will overturn using 21 input features related to earthquakes and structure, providing decision support for seismic design in structural engineering. The project is developed and maintained by Qingle Cheng, open-sourced on GitHub, and the related paper is currently under review.

Earthquakes are major natural disasters threatening lives and property. Accurately predicting structural seismic responses is crucial for design and disaster prevention. Traditional methods (finite element simulation, mechanical models) face challenges in handling complex structures and nonlinear behaviors. Rocking structures may rotate rigidly during earthquakes, having self-resetting capabilities but also the risk of overturning. Predicting whether they will overturn is an important topic. This project proposes a data-driven method to quickly predict overturning risk by learning the relationship between historical data and responses.

Model Architecture: Uses deep neural networks (DNN) to leverage their nonlinear fitting ability to learn complex mappings between features and results. Task Definition: Binary classification task, outputting overturning probability (0-1) and binary result (overturn/non-overturn). Input Features: A total of 21 features, covering seismic ground motion characteristics (17 categories with 20 items: PGA, PGV, PGD, PGA/PGV, ARMS, VRMS, DRMS, CAV, CAD, CAI, significant duration, bracketed duration, PSA, PSV, Ic, ASI, VSI, Tp, Arias intensity, MIV) and structural features (number of floors), reflecting professional knowledge in seismic engineering.

Development Environment: Python 3.12.6, dependent on libraries like PyTorch, NumPy, Pandas, scikit-learn, supporting CPU/GPU operation. Model Files: Provides pre-trained model best_model_300eps.pt and scaler scaler.pkl. Inference Interfaces: Supports single-sample prediction (input 21 feature values) and batch prediction (CSV file input/output).

Application Scenarios: Seismic design assistance (optimizing schemes), seismic risk assessment (safety of existing buildings), emergency response decision-making (post-earthquake damage assessment), research tool (analyzing factor impacts). Data Foundation: Training data comes from the PEER NGA strong motion database; the author thanks the database for providing high-quality seismic ground motion records.

Technical Advantages: Data-driven (no need for complex mechanical models), fast prediction (millisecond-level inference), probabilistic output (provides decision information), specialized features (covers key characteristics). Limitations and Improvements: Dependent on training data quality; DNN black-box nature (needs explainable AI); only applicable to rocking rigid bodies (expand structure types); requires stricter uncertainty quantification (e.g., Bayesian neural networks).

This project demonstrates the application value of deep learning in seismic engineering, providing practical pre-trained models and tools with professionalism and scalability. In the future, combining data-driven approaches with physical knowledge is expected to solve more complex engineering problems, improve the efficiency of seismic design and risk assessment, and contribute to disaster prevention and mitigation.