The article "Transformer model for sensitivity analysis of steel and steel-fiber reinforced concrete beams", written by Stefanie Schoen, Steffen Freitag, Vladislav Gudzulic, and Günther Meschke, has been published in "Advances in Engineering Software" by Elsevier.

Abstract:
Due to inherent uncertainties, it is essential to quantify both aleatory and epistemic uncertainties when assessing the structural behavior and reliability of reinforced concrete (RC) and steel-fiber reinforced concrete (SFRC) structures, as these uncertainties can significantly impact load-bearing capacity and crack development. To enable fast predictions during the design process, circumventing time consuming finite element simulations, and considering implicitly material and structural uncertainties, a novel Transformer-based surrogate model is proposed in this paper. The surrogate model efficiently predicts the history-dependent response of RC and hybrid RC-SFRC beams, specifically, load–displacement and maximum crack width-displacement curves. Unlike conventional feedforward neural networks, the Transformers captures long-range dependencies across the entire loading process in parallel, making it well-suited for path-dependent structural behavior. To assess the influence of key uncertainties, the surrogate model is applied within a systematic sensitivity analysis. Results show that the concrete cover dominates the influence on the load–displacement behavior in RC beams, while the fiber properties govern the response in hybrid RC-SFRC beams. The findings demonstrate the potential of Transformer models as a computationally efficient tool for reliability assessment in structural engineering.

Before May 21, 2026 this share link provides a 50 days' free access to the article:
https://authors.elsevier.com/c/1msaE3Rf7bHSTK.

Alternatively, try this link to gain access:

Chen Xu presented his doctoral theses with the title "Physics Meets Data: Machine Learning Frameworks for Structural Engineering" on 31st March 2026 at 1:30PM.

Xu.jpg" width="500" height="600">
Dear Chen, well done, and congratulations! We wish you all the best for your future, both professionally and personally.


Abstract:
Structural engineering increasingly requires models that are both data-driven and physically consistent, especially when real-world measurements are sparse, noisy, and heterogeneous. This thesis develops physics-informed machine learning (PIML) frameworks that combine physical knowledge with observations to address three key challenges. First, it enables the identification of unknown loads in tunnel linings from limited displacement measurements. Second, it supports real-time prediction of tunnelinginduced ground settlement by integrating simulations with monitoring data. Third, it improves concrete damage classification in structural health monitoring by transferring knowledge from synthetic data to real experiments. Across these applications, the proposed approaches enhance accuracy, efficiency, and robustness, demonstrating the potential of PIML for more reliable structural analysis, monitoring, and maintenance.

The lecture dates for the summer term 2026 are online:

"Adaptive insertion of interface elements for fracture analysis: Reliable computation of interface traction" is the title of the recent open access publication by Koussay Daadouch, Vladislav Gudzulic, and Günther Meschke. It has been published by Elsevier in the journal "Computer Methods in Applied Mechanics and Engineering".

Abstract:
The cohesive zone model using interface elements is a practical and widely adopted approach for modeling crack initiation and propagation. However, enriching a model with interface elements significantly increases computational costs due to node duplication. To mitigate this, numerous adaptive insertion strategies have been developed to insert interface elements on the fly only when and where needed. Existing strategies rely on stress-based insertion criteria, which often fail to ensure timely and accurate placement of interface elements. Moreover, many existing approaches suffer from a critical limitation: poor configurations of inserted interface elements lead to significant errors in traction computation. In this paper, we investigate the state-of-the-art adaptive insertion methods focusing on the influence of interface elements configurations on traction accuracy. Based upon the findings we propose a novel algorithm that reliably computes the traction of interface elements and serves as a robust and precise insertion criterion, alleviating the limitations of existing techniques. The algorithm leverages the unique formulation of linear interface elements, enabling traction evaluation in an efficient post-processing step without requiring node duplication. Finally, we present a numerical simulation campaign that highlights the error trends inherent to existing adaptive insertion schemes and demonstrates the efficacy of the proposed method.

Chen Xu, Giao Vu, Ba Trung Cao, Zhen Liu, Fabian Diewald, Yong Yuan, and Günther Meschke are the authors of the newly published article entitled "Bridging simulation and experiment: A self-supervised domain adaptation framework for concrete damage classification". It has been published by Elsevier in the journal Advanced Engineering Informatics.

Abstract:
Reliable assessment of concrete degradation is critical for ensuring the safety and longevity of engineering structures. This study proposes a self-supervised domain adaptation framework for robust concrete damage classification using coda wave signals. To support this framework, an advanced virtual testing platform is developed, which combines multiscale modeling of concrete degradation with ultrasonic wave propagation simulations. This setup enables the generation of large-scale labeled synthetic data under controlled conditions, reducing the dependency on costly and time-consuming experimental labeling. However, neural networks trained solely on simulated data often suffer from degraded performance when applied to experimental data due to domain shifts. To bridge this domain gap, the proposed framework integrates domain adversarial training, minimum class confusion loss, and the Bootstrap Your Own Latent (BYOL) strategy. These components work jointly to facilitate effective knowledge transfer from the labeled simulation domain to the unlabeled experimental domain, leading to accurate and reliable damage classification in concrete. Extensive experiments demonstrate that the proposed method achieves notable performance gains, reaching an accuracy of 0.7762 and a macro F1 score of 0.7713, outperforming both the plain 1D CNN baseline (accuracy: 0.5867; macro F1: 0.5832) as well as six representative domain adaptation techniques. Moreover, the method exhibits high robustness across independent runs and adds only minimal training overhead (about two additional minutes). These findings underscore the practical potential of the proposed simulation-driven and label-efficient framework for real-world structural health monitoring applications.

Get a free access with this authors' share link (until January 12, 2026) or try this link: