Conference Agenda

Reliable prediction of fatigue lifetime is essential for ensuring the durability of structures subjected to cyclic loading, particularly when non-uniform load histories are involved. Such scenarios pose a significant challenge, as load sequence effects strongly influence fatigue behavior while making conventional highcycle fatigue simulations computationally demanding. This study introduces a physics-based machine learning (ϕML) framework for efficient prediction of fatigue lifetime under variable loading conditions. The approach employs a feedforward neural network in which experimentally observed fatigue characteristics are embedded as physical constraints within a customized loss function, enhancing generalization and physical consistency. The network is trained using simulation data generated by an anisotropic continuum damage model for fatigue. Model calibration and validation are performed using experimental fatigue data of concrete cylinders under uniaxial compression, with the training data capturing the effect of different loading sequences. The ϕML model shows superior performance compared to purely data-driven neural networks, particularly when only limited data are available. A general prediction algorithm based on the trained ϕML model is then applied to complex loading scenarios involving multiple load transitions. The obtained results reproduce experimentally observed reductions in fatigue lifetime with increasing load jumps, demonstrating the model’s robustness and interpretability. Overall, the proposed ϕML approach offers a computationally efficient and physically consistent framework for fatigue life prediction and holds strong potential for integration into digital twin systems for real-time structural health monitoring.

Addressing age-related deterioration of road bridges requires efficient crack detection methods to replace traditional visual inspections. This study proposes a deep learning-based approach for detecting pavement cracks in RC bridges using UAV-captured visible images. The method addresses the challenge of varying illumination conditions by implementing a two-stage process: (1) scene classification to distinguish noisy areas, sunlit pavement areas, and shaded pavement areas, and (2) specialized crack detection models trained for specific lighting conditions. U-Net architecture is employed for both scene classification and crack detection models. Image processing techniques are applied in L*a*b* color space to enhance model performance. The approach is validated on three in-service RC bridges. Results demonstrate high-precision scene classification and improved crack detection accuracy when using lighting condition-specific models. The highest detection accuracy is achieved when the lighting conditions used for model training match those of the target detection areas, confirming the effectiveness of condition-specific training approaches.

This study compares LSTM neural networks and physics-based heat balance models for predicting concrete dam surface temperatures under shadow effects from surrounding terrain. Three-dimensional shadow modeling is integrated with UAV-LiDAR point cloud data to quantify shadow-induced radiation variations. Monitoring data from 2023-2024 are analyzed to evaluate physical consistency and spatial generalization capability. Strong correlation (R²=0.562-0.587) between radiation reduction and temperature changes is demonstrated by the heat balance model. In contrast, near-zero correlation (R²=0.0001-0.0083) is observed in LSTM predictions. Spatial analysis reveals identical predictions (r=0.9999) across different monitoring locations by LSTM, indicating that spurious temporal correlations are learned rather than causal shadow mechanisms. These findings demonstrate that high accuracy metrics cannot guarantee physical correctness in data-driven models. Physics-based or hybrid approaches are recommended for structural health monitoring applications requiring spatial generalization beyond train-ing data locations.

Projectile penetration into multi-layer composite targets involves highly non-linear and dynamic physical processes, including stress wave propagation, material failure, large deformations, and interactions between different material layers. These phenomena are governed by partial differential equations (PDEs) representing conservation laws and constitutive behaviour under high strain-rate conditions. Solving these equations accurately is essential for understanding and predicting impact behaviour. However, conventional numerical methods such as finite element (FE) approaches, though accurate, can be computationally expensive and timeconsuming. This limitation becomes critical when performing multiple simulations across input parameters such as impact velocity, material properties, projectile geometry, and boundary conditions. This work introduces a Physics-Informed Neural Network (PINN) framework to model projectile penetration in multi-layer composite structures. Unlike traditional data-driven machine learning models that require large amounts of labelled data, PINNs embed the physical laws of the problem directly into the learning process. The governing PDEs and relevant initial and boundary conditions are incorporated into the loss function. This allows the network to learn physically consistent solutions without relying on labelled simulation or experimental data, making the method particularly useful when data collection is limited or computationally costly. Physics-based feature engineering is applied to improve the PINN’s learning capability. Features informed by impact mechanics are inputs to the neural network. These physically meaningful inputs help the network better understand the structure of the solution space and enhance generalization across a range of impact scenarios. The model is trained to satisfy all relevant physical constraints, ensuring that the solution respects the underlying mechanics of high-velocity penetration events. An adaptive loss balancing technique is employed to improve the stability and efficiency of the training process. During training, the different components of the loss function—representing the PDEs, boundary conditions, and initial conditions—may evolve at different rates, leading to imbalance and slower convergence. The adaptive approach dynamically adjusts the weighting of each component, ensuring that all physical constraints are learned effectively and in proportion. This technique improves convergence behaviour and leads to more accurate and robust solutions. The proposed PINN framework is validated with the experimental results under various impact conditions. These include different projectile velocities, shapes, and layered target configurations. The results show that the PINN can accurately predict key physical quantities such as the depth of penetration. The model strongly agrees with experimental results and generalizes unseen scenarios well. Once trained, the model can deliver real-time predictions, offering a fast and efficient alternative to traditional simulation methods for repeated evaluations or parametric studies. Integrating physical knowledge into the neural network architecture makes the PINN a promising tool for simulating complex impact problems computationally efficiently. This approach offers significant advantages in the study and design of advanced protective systems across a range of engineering applications, including defence, aerospace, and civil infrastructure.