2026-02-20

The paper "Multi-task deep learning-based failure diagnosis for tubular joints: Automated identification and description generation" was published in the journal Advanced Engineering Informatics.

This study addresses the critical yet vulnerable tubular joints in offshore platforms and bridges, proposing a multi-task deep learning model for simultaneous failure identification and description generation. Traditional failure inspection methods are time-consuming and labor-intensive, while existing intelligent methods often focus on single-task or single-modal designs, which fail to fully leverage multimodal data and struggle to capture coexisting failure modes, limiting comprehensive diagnostic assessments. To overcome these challenges, a multimodal database is constructed from 141 experimental programs, including 409 failed joints with images, textual descriptions, and component-level labels, further augmented to 1,227 images and 3,681 sentences. The proposed model employs a pretrained encoder for image feature extraction, a multi-label classification decoder for failure identification, and an attention-based image captioning decoder for failure description generation. Experimental results demonstrate that the model achieves a BLEU-4 score of 76.70 and an mAP of 0.9467, outperforming single-task baselines. Validation on real engineering images further shows its preliminary transferability, highlighting the potential of the multi-task learning framework for automated failure diagnosis.

Note: Advanced Engineering Informatics is a top-tier Q1 journal in the field of engineering technology, with an impact factor of 9.9 in 2025. The first author of the paper is Zhang Wenhao, a PhD student at Southern University of Science and Technology, with Professor Hou Chao as the corresponding author. The research was funded by the National Natural Science Foundation of China (NSFC) , the National Key Research and Development Program of China, and Shenzhen Science and Technology Program.

2026-02-12

The paper titled “Deep learning-based real-time multiload response prediction and inverse analysis of offshore bridges” has been published in Engineering Structures.

Offshore bridges operate in complex ocean environments, making structural analysis, design, and monitoring more challenging. Conventional time-history analysis and nonlinear model updating based on finite-element methods are computationally intensive and time-consuming, which limits their use in many scenarios. To develop a more efficient analytical tool, this study proposes a deep learning-based offshore bridge predictor, DeepOBP. By integrating structural characteristics with coupled dynamic loads in ocean environments, DeepOBP enables millisecond-level, high-precision nonlinear dynamic offshore bridge response predictions. Building on this surrogate model, a differentiable structural inverse framework (Inverse DeepOBP) is further developed, coupling the predictor with gradient-based optimization to support rapid damage identification and model calibration for structural health monitoring.

Experimental results show that DeepOBP achieves high accuracy under both normal operating conditions and multi-hazard coupled conditions, with R² = 0.93 and 0.92, respectively. Inverse DeepOBP delivers more than a 10-fold speed-up over surrogate-based model updating using a heuristic algorithm, and more than a 10⁴-fold speed-up over nonlinear finite-element model updating, while maintaining relative errors below 7% for each identified parameter. These results demonstrate the potential of the proposed framework for efficient structural analysis and real-time monitoring of offshore bridges.

Note: Engineering Structures is a Q1 journal in the Engineering and Technology field, with a 2025 impact factor of 6.4. The first author of the paper is Master’s student Zhuyu Sun, and Prof. Yutao Guo is the corresponding author. This work was supported by the National Key Research and Development Program of China and the Shenzhen Science and Technology Program.

2026-02-11

The paper entitled “ROM-PINN: A physics-informed neural network with reduced-order modelling for nonlinear structural response prediction” has been published in the journal Structures.

To address the high computational cost of conventional finite element time-history analysis for nonlinear structural responses and the lack of physical consistency in purely data-driven neural networks, this study proposes a Reduced-Order Model Physics-Informed Neural Network (ROM-PINN). The proposed method embeds the nonlinear reduced-order dynamic equilibrium as a physics constraint in network training, while learning nonlinear effects through a virtual-force representation. By integrating reduced-order projection with a sensor-aware deployment strategy, ROM-PINN enables efficient and physically consistent response prediction for high-dimensional structural systems from limited observations. A case study on an offshore bridge demonstrates that ROM-PINN achieves accurate multi-sensor response prediction and provides clear quantitative improvements over a purely data-driven baseline: the mean squared error (MSE) decreases by 10.0%, the physics-consistency metric (PR) increases by 83.7%, and under 10% measurement noise the MSE is further reduced by 23.3%. These results highlight the strong potential of ROM-PINN for practical engineering applications such as rapid analysis and online monitoring of large-scale nonlinear structures.

Note: Structures is a Q1 journal in the Engineering and Technology category, with an Impact Factor of 4.3 (2025). The first author of the paper is Master’s student Zhu-Yu Sun, and Dr. Yu-Tao Guo served as the corresponding author. This work was supported by the National Key Research and Development Program of China and the Shenzhen Science and Technology Program.

2026-01-22

The paper "Multiscale investigation of ternary precursor proportioning in engineered geopolymer composites: Effects of silica fume replacement ratio and GGBS content" has been published in the journal Construction and Building Materials. 

To reduce cement consumption and promote eco-friendly construction materials, engineered geopolymer composites (EGC) have been developed as sustainable alternatives to engineered cementitious composites (ECC). This study systematically investigates the effects of precursor tailoring, specifically silica fume (SF) replacement ratio (5–15 %) and ground granulated blast-furnace slag (GGBS) content (20–80 %), on the mechanical properties, microstructure, and sustainability of EGC. The results show that appropriate precursor composition enables a wide strength coverage, with compressive strengths ranging from 48 to 117 MPa, meeting the requirements of different structural applications. Digital image correlation (DIC) and in-situ crack analysis reveal stable multiple cracking behavior, with ultimate tensile strains consistently exceeding 8 % and effective crack width control within 120 μm. These findings confirm the achievement of multiple cracking and high tensile ductility through coordinated fibre/matrix interaction and tailored precursor design. Additionally, microstructural characterizations (XRD, FTIR, TGA) show that higher GGBS contents promote C-(N)-A-S-H gel formation and a degree of geopolymerization, leading to matrix densification and enhanced mechanical properties. Moderate SF replacement also contributes to pore filling, further increasing the matrix compactness. From a sustainability perspective, a cradle-to-gate life cycle assessment (LCA), using typical M45-ECC as a benchmark, shows that the developed EGC achieve approximately 50 % lower embodied carbon while maintaining comparable manufacturing feasibility. Overall, this study demonstrates that precursor-tailored EGC can simultaneously satisfy structural performance criteria and sustainability targets, highlighting their potential for low-carbon practical applications.

Note: Construction and Building Materials is a top journal in the field of engineering and technology in Q1, with an impact factor of 8.0 in 2025. The first author of the paper is doctoral student Wan Feihong, and Professor Guo Yutao is the corresponding author. The research results were funded by the Shenzhen Science and Technology Program and the National Natural Science Foundation of China.

2026-01-10

The paper "Low-carbon remediation of contaminated marine mud sediment for efficient in-situ recycling and application" has been published in the journal ENGINEERING Environment. 

This study addresses the issue of difficult disposal of contaminated marine sediment and its constraints on waste management by proposing a low-carbon solidification treatment using aluminosilicate raw materials to achieve in-situ resource utilization as engineering backfill material. By incorporating a mixture of 25% ordinary Portland cement, fly ash, slag, and 5% river sand, the unconfined compressive strength of the solidified body can reach up to 8.69 MPa, and it can efficiently stabilize heavy metals, making the product comply with environmental safety standards in both China and the United States. XRD analysis reveals that the material is mainly composed of SiO₂, with the formation of secondary phases such as calcium carbonate, iron-manganese oxides, and complex silicates, which collectively endow it with excellent structural mechanical properties. This technology not only provides a feasible path for the large-scale recycling and utilization of contaminated marine silt but also helps improve resource efficiency, protect the environment, and support the realization of carbon reduction and carbon neutrality goals. 

Note: ENGINEERING Environment, formerly known as Frontiers of Environmental Science & Engineering, is a journal in the Q1 zone of the field of environmental science and ecology, with an impact factor of 6.4 in 2025. Dr. Dassekpo is the corresponding author. The research results were supported by the  Guangdong  Basic and  Applied  Basic  Research  Foundation.