Physics-guided Convolutional Neural Network (PhyCNN) for Data-driven Seismic Response Modeling (1909.08118v1)

Abstract: Seismic events, among many other natural hazards, reduce due functionality and exacerbate vulnerability of in-service buildings. Accurate modeling and prediction of building's response subjected to earthquakes makes possible to evaluate building performance. To this end, we leverage the recent advances in deep learning and develop a physics-guided convolutional neural network (PhyCNN) framework for data-driven seismic response modeling and serviceability assessment of buildings. The proposed PhyCNN approach is capable of accurately predicting building's seismic response in a data-driven fashion without the need of a physics-based analytical/numerical model. The basic concept is to train a deep PhyCNN model based on available seismic input-output datasets (e.g., from simulation or sensing) and physics constraints. The trained PhyCNN can then used as a surrogate model for structural seismic response prediction. Available physics (e.g., the law of dynamics) can provide constraints to the network outputs, alleviate overfitting issues, reduce the need of big training datasets, and thus improve the robustness of the trained model for more reliable prediction. The trained surrogate model is then utilized for fragility analysis given certain limit state criteria (e.g., the serviceability state). In addition, an unsupervised learning algorithm based on K-means clustering is also proposed to partition the limited number of datasets to training, validation and prediction categories, so as to maximize the use of limited datasets. The performance of the proposed approach is demonstrated through three case studies including both numerical and experimental examples. Convincing results illustrate that the proposed PhyCNN paradigm outperforms conventional pure data-based neural networks.

• The paper introduces Physics-guided Convolutional Neural Network (PhyCNN), which integrates physics principles into CNNs to improve seismic response modeling with limited data.
• Numerical and experimental results show PhyCNN significantly enhances predictive accuracy, achieving a correlation coefficient of 0.95 in displacement prediction compared to 0.60 with standard CNNs.
• PhyCNN bridges data-driven and physics-based models, offering a robust and adaptable solution for structural assessments, especially in scenarios with limited sensor data.

Physics-guided Convolutional Neural Network (PhyCNN) for Data-driven Seismic Response Modeling

The paper by Zhang, Liu, and Sun presents a method called Physics-guided Convolutional Neural Network (PhyCNN), aimed at enhancing data-driven seismic response modeling of structures. The approach leverages convolutional neural networks (CNNs) integrated with physics knowledge to accurately predict structural responses to seismic activities using limited data, thus addressing limitations inherent in traditional data-driven models in structural dynamics.

Summary and Key Findings

The proposed PhyCNN model is formulated to incorporate principles of physics, such as the laws of dynamics, into the neural network training process. This integration serves multiple purposes:

1. It establishes constraints for the network's outputs, improving the prediction accuracy by mitigating overfitting issues.
2. It reduces the dependence on large training datasets, enhancing the model's robustness and applicability.

The paper employs an unsupervised learning algorithm based on K-means clustering to maximize the utility of limited datasets. This algorithm effectively partitions data into training, validation, and prediction categories.

The authors highlight the PhyCNN's capabilities by benchmarking its performance against traditional CNNs devoid of physics-guidance. The results demonstrate superior predictive accuracy of the PhyCNN model in forecasting building seismic responses without relying on purely physics-based analytical models. Such enhanced performance is crucial for accurate structural serviceability assessments and fragility analyses.

Numerical Results and Experimental Validation

Quantitative assessment of the PhyCNN indicates significant improvements in predictive accuracy across several metrics. For instance, the inclusion of physics constraints notably increases the correlation coefficients (r) in predictive tasks when compared to non-physics-guided counterparts.

1. Numerical Examples: The numerical validation involved two nonlinear structural scenarios. Noteworthy is the first example where the correlation coefficient for displacement prediction using PhyCNN reached r = 0.95, compared to 0.60 with the unenhanced CNN. This affirms the advantage of embedding physics constraints.
2. Experimental Example: In a real-world application, the PhyCNN was trained using data from a 6-story building in San Bernardino, CA. The model efficiently predicted structural displacements during unseen seismic events with low error margins. Error distribution analysis affirmed a prediction accuracy within a 5% threshold for most cases, providing empirical support for the model's practical applicability.

Implications and Future Directions

The integration of physics-based constraints in neural network architectures like PhyCNN sets a precedent for advancing predictive modeling efficacy in seismic engineering. This methodology bridges a critical gap between data-driven approaches and physics-based models, making it adaptable to scenarios with limited sensor data availability. Its application can extend beyond seismic response modeling to other structural assessments and hazard types.

Future developments may involve extending the PhyCNN framework to incorporate real-time adaptive learning, further enhancing its scalability and applicability to diverse structural and environmental contexts. Additionally, exploring augmented reality (AR) or virtual reality (VR) interfaces for monitoring and interpreting seismic data in conjunction with PhyCNN predictions could revolutionize structural health monitoring systems.

In summary, the PhyCNN model represents a significant advancement in data-driven seismic response modeling, offering a robust and adaptable solution that effectively utilizes limited data inputs while incorporating essential physical laws for heightened predictive reliability.