Rapid post-event assessment of earthquake damage is essential for resilient emergency response and risk mitigation. We present a multi-scenario deep learning framework that uses stacked LSTM and a hybrid LSTM–RNN to (i) forecast structural response variables (displacement \(u\) , velocity \(v\) , acceleration \(a\) , and Damage Index \(DI\) ), (ii) classify damage status, (iii) conditionally estimate the weight factor \(w\) for damaged cases ( \(DI\ge 1\) ), and (iv) identify features most associated with negligible damage via GridSearchCV. In addition, we benchmark a quantum-inspired Activation-based Probabilistic Machine (APM) classifier head attached to the shared sequence encoder to probe whether compact state encodings and Hadamard-style interactions can improve damage discriminability under the same leakage-safe pipeline. Models were trained on 40-step windows for 100 epochs with Adam, using linear heads for regression and Softmax for classification (APM and standard heads share the global training schedule). Across sequence-regression tasks, the LSTM–RNN consistently outperformed stacked LSTM: \({R}^{2}\) improved from 98.37 to 99.42% for \(u\) , 89.57 to 97.58% for \(v\) , 97.69 to 99.8% for \(a\) , and 99.68 to 99.97% for \(DI\) . For \(DI\) , error metrics were markedly lower with LSTM–RNN (MAE 0.0031 vs. 0.0132, RMSE 0.0047 vs. 0.0163, MAPE 1.51 vs. 5.25, MedAE 0.0022 vs. 0.0118), indicating tighter tracking of the damage signal. The conditional \(w\) estimation and feature-ranking scenario offers practical levers for risk-informed prioritization, while the APM head provides a compact, quantum-inspired alternative for damage classification within the same framework. Overall, the results support sequence models, particularly LSTM-RNNs, as an effective basis for rapid, data-driven earthquake damage modeling and decision support, with quantum-inspired heads as a complementary option.
