Understanding and replicating driver lane-changing behavior is essential for developing advanced driver assistance systems that align with driver expectations and for constructing dynamic, realistic traffic flow scenarios in autonomous driving tests. This study proposes a two-dimensional, force-based lane-changing dynamics model that represents vehicle motion as acceleration changes under the influence of (surrounding) social forces. The model captures longitudinal and lateral interactions with neighboring vehicles and incorporates personalized parameters to reflect trajectory-level heterogeneity. A temporal attention mechanism is introduced to model the dynamic shift of driver focus during lane change, ensuring continuity in interaction forces. For parameter calibration, AdaP+ is developed, a gradient-based optimizer that integrates Nesterov momentum, decoupled weight decay, and gradient belief within the Adam framework, and improved with a pruning strategy for better initialization and faster convergence. Experimental findings on the HighD dataset demonstrate that the social force lane change model outperforms baselines in both trajectory fitting and behavioral simulation. Gradient-based optimizers, particularly AdaP+, achieve higher precision than swarm intelligence algorithms. Overall, the proposed model accurately replicates complex lane-change dynamics, while the integration of personalized parameters and attention mechanism effectively captures trajectory-level heterogeneity and interaction continuity, with potential to support naturalistic, personalized lane-change trajectory generation in future connected-vehicle settings.

Rapid and accurate prediction of structural response histories across multiple seismic hazard levels is essential for seismic performance and resilience evaluation. While surrogate modeling for peak values of individual responses has emerged as a popular solution, existing approaches often struggle to predict time-history-related quantities, particularly permanent residual displacements, due to highly nonlinear hysteretic behaviors during earthquakes. To address this, this study presents a surrogate model development framework for predicting seismic response time-histories using a physics-informed temporal convolutional network integrated with bidirectional gated recurrent units, referred to as PhyTCN-BiGRU . Through enforcing physical consistency through equation-of-motion constraints, the proposed framework effectively captures complex path-dependent dynamics. Validated on a Bouc-Wen system and a nonlinear four-story steel frame, the model demonstrates superior accuracy in predicting full time-history responses, including peak and residual demands, under varying seismic intensities. PhyTCN-BiGRU demonstrates strong capability in reproducing nonlinear seismic response time-histories under earthquakes with both low and high intensities. Critical engineering demand parameters (EDPs) are predicted with exceptional accuracy. The results confirm that PhyTCN-BiGRU offers a reliable and computationally efficient surrogate for seismic response history prediction, including peak and residual values, which can be further used for performance-based seismic assessment and design.

Accurate prediction of TBM (Tunnel Boring Machine) operational parameters is critical to ensure safe and efficient tunnel construction. However, traditional high-accuracy models often face scalability issues due to excessive parameter counts which limit their applicability for real-time updates with on-site data streams. To address this, we propose a novel cutterhead thrust prediction framework integrating clustering-based transfer learning and optimized data augmentation strategy. The three-tiered approach includes: (1) clustering raw thrust data into sub-databases by analyzing excavation cycles' variation forms; (2) a transfer learning classifier is trained to recognize the variation form of incoming data and route samples to specialized predictors, within which multiple candidate models are trained and the best-performing expert is selected for deployment; (3) employing LAFPI (Leading Averaged Force-Penetration Index) for risk warning of abrupt surrounding rock type changes during TBM excavation. Experimental results indicate that the proposed cutterhead thrust prediction framework attains high accuracy (R² > 0.85) over prediction horizons exceeding 1,000 time steps, while deploying only a 0.015B-parameter expert model at inference and requiring about 2 minutes of training. Compared with the smallest conventional baseline models that can stably fit the dataset (LSTM/GRU/Transformer), the deployed expert model achieves an approximately 95% reduction in static parameters and an approximately 85% reduction in per-epoch training time. This establishes its superiority in predictive accuracy, computational efficiency, and adaptability for real-time TBM monitoring systems, enabling fine-tuning the model using on-site data streams.

Stochastic Hybrid Simulation (SHS) requires reliable prediction of structural responses to quantify uncertainties, yet conducting repeated runs via nonlinear time-history analysis or shaking-table testing is computationally or experimentally prohibitive. Although neural networks show promise for nonlinear response prediction, they struggle under the extremely limited high-fidelity data typical of experimental substructures in SHS. While Physics-Informed Neural Networks (PINNs) reduce data demands, they strictly rely on governing equations, which often fail to capture the complex, unmodeled nonlinearities (e.g., friction, noise) inherent in real physical specimens. Consequently, existing methods face challenges in accurately modeling experimental substructures. In this study, a Prior-Knowledge-Infused Neural Network (PKNN) is developed as a high-fidelity surrogate to replace the physical substructure in SHS. A two-stage training strategy is proposed: the network is first pretrained using abundant data from a simplified numerical model to learn fundamental physical laws, and subsequently fine-tuned with scarce shaking-table measurements to correct the physics-based bias and capture complex realistic behaviors. A 1/3-scale steel frame is used to evaluate the approach, predicting maximum inter-story drift ratios under varying structural parameters and ground-motion intensities. Results indicate that, in data-scarce regimes, PKNN maintains significantly higher predictive stability and accuracy than conventional neural networks or polynomial chaos expansions. These findings suggest that embedding simplified physics through pretraining offers a practical means to reduce the demand for high-fidelity data while retaining the ability to model nonlinear response and offers a robust pathway for SHS-oriented experimental design and uncertainty quantification.