Rapid earthquake damage assessment via hybrid LSTM-RNN with a quantum-inspired classification head based on...

Area of Science:

• Earthquake Engineering
• Artificial Intelligence
• Structural Health Monitoring

Background:

• Rapid post-earthquake damage assessment is crucial for effective emergency response and risk mitigation.
• Accurate forecasting of structural response variables and damage classification are key challenges.

Purpose of the Study:

• To develop and evaluate a deep learning framework for rapid earthquake damage assessment.
• To compare the performance of stacked LSTM and LSTM-RNN models.
• To benchmark a quantum-inspired Activation-based Probabilistic Machine (APM) classifier.

Main Methods:

• A multi-scenario deep learning framework utilizing stacked LSTM and LSTM-RNN architectures.
• Forecasting structural response variables (displacement, velocity, acceleration, Damage Index) and classifying damage status.
• Benchmarking a quantum-inspired APM classifier head for damage discriminability.

Main Results:

• The LSTM-RNN model consistently outperformed stacked LSTM across all sequence-regression tasks, showing significant improvements in accuracy and reduced error metrics.
• The APM head demonstrated potential as a compact, quantum-inspired alternative for damage classification.
• The framework successfully estimated conditional weight factors and identified features associated with negligible damage.

Conclusions:

• Sequence models, particularly LSTM-RNNs, provide an effective foundation for data-driven earthquake damage modeling and decision support.
• The developed framework offers practical tools for risk-informed prioritization in post-earthquake scenarios.
• Quantum-inspired approaches present a promising complementary option for enhancing damage classification capabilities.