Research Group of Xinzheng Lu, Dept. Civil Engineering, Tsinghua University, Beijing (2026)

19/08/2026

Our new paper “QuakeF2G: A region-specific transformer-based ground motion prediction model considering fault segments modeling” has been published in Computer-Aided Civil and Infrastructure Engineering

Accurately capturing near-fault ground motions and rupture directivity remains challenging. Empirical ground-motion models often simplify complex rupture geometry, while physics-based simulations can be computationally demanding.

In this work, we developed QuakeF2G, a Transformer-based probabilistic ground-motion prediction framework that represents earthquake ruptures as a sequence of fault-segment tokens. A dedicated fault encoder and geometry-aware cross-attention mechanism integrate rupture geometry, source–site configuration, and site information to predict PGV.

Some results from the study:

• Trained on approximately 626,000 CyberShake earthquake scenarios in Southern California
• Explicit finite-fault representation reduced RMSE by 36% compared with the point-source baseline
• The same framework supports both ground-motion prediction and conditional spatial interpolation
• In our spatial interpolation experiments, QuakeF2G outperformed Ordinary Kriging across the evaluated sampling densities

Code: https://github.com/YitianF/QuakeF2G
Paper: https://doi.org/10.1016/j.cacaie.2026.100183

Many thanks to Yitian Feng and Weiqiang Zhu for the collaboration.

The graphical abstract accompanying this post was generated using AI.

14/08/2026

Congratulations to Yongjun Kang on the publication of the review paper, “Advances and Applications of Computational Design Methods in Structural Design,” in the ASCE Journal of Computing in Civil Engineering.

Computational design is reshaping structural engineering, but a clear framework is still needed to connect digital representation, performance analysis, and design generation. This paper organizes the field into three complementary categories:

• Drawing-oriented methods, including CAD and BIM
• Analysis-oriented methods, including formula-based calculations and finite-element analysis
• Design-oriented methods, including form-finding, structural optimization, and generative structural design

The paper reviews their historical development, engineering applications, current limitations, and future directions. A key message is that these three paradigms are not successive replacements for one another; rather, they need to work together across the structural design workflow.

Future progress will depend on stronger semantic interaction, more robust engineering reasoning and modeling, higher-quality data, and more accurate and interpretable AI-generated designs.

Paper: https://doi.org/10.1061/JCCEE5.CPENG-7548

The graphical abstract accompanying this post was generated using AI.

12/08/2026

Our new paper, “Graph-based data-physics hybrid surrogate modeling for seismic response estimation of steel moment frame building portfolios,” is now online in Bulletin of Earthquake Engineering.

Regional-scale seismic assessment requires models that are both accurate and computationally efficient. In this study, we represent steel moment-frame buildings as graphs and integrate graph neural networks with physics-based multi-degree-of-freedom models.

The proposed framework:

• preserves building topology and member-level structural information;
• accommodates buildings with different heights and configurations without building-specific retraining;
• reduces mean absolute error by up to 57.5% compared with conventional parameterized MDOF models; and
• enables response estimation in less than one second—or about 0.007 seconds using the end-to-end model.

The hybrid approach also better preserves physical response trends at high drift levels and under critical seismic demands, providing a promising pathway for scalable seismic assessment based on urban building databases and digital twins.

The datasets and GNN algorithms are available at:
Dataset: https://doi.org/10.57760/sciencedb.33049
Code: https://doi.org/10.57760/sciencedb.33051

Read the paper: https://doi.org/10.1007/s10518-026-02629-z

The graphical abstract was created with the assistance of generative AI.

10/08/2026

Young Scholars Workshop on Intelligent Engineering Design

Today we held the Young Scholars Workshop on Intelligent Engineering Design at Tsinghua University. The workshop featured 27 young speakers from 16 universities across mainland China, Hong Kong, and overseas—with 23 presenting onsite and four online—and attracted an audience of more than 60 participants.

The presentations covered a wide range of topics, including data–physics-driven surrogate models, graph neural networks, topology and differentiable optimization, generative AI for architectural and structural design, BIM automation, and LLM-based design agents.

Two impressions stood out.
First, AI-enabled engineering design is advancing at remarkable speed, moving rapidly from individual algorithms toward integrated workflows for design, analysis, checking, and optimization.
Second, the creativity of these young researchers was equally striking. The presentations were technically rigorous, highly original, and genuinely exciting, with many ideas closely connected to real engineering challenges.

Many thanks to all the speakers and participants for the excellent presentations and open discussions. I look forward to seeing these ideas develop into further research and collaboration.

30/07/2026

Our latest paper, “An intelligent sizing design method for frame-shear wall components in industrial parks via heterogeneous graph neural networks,” has been published in Advanced Engineering Informatics.

Our group has been continuously advancing AI-assisted structural design. While previous studies have achieved significant progress in structural layout generation, layout alone does not constitute a complete structural design—component sizing is the essential bridge from layout generation to structural analysis and engineering implementation. Existing intelligent sizing methods often focus on pure frame or shear-wall systems, cover only selected component types, and rely mainly on geometry and global design conditions without explicitly representing component-level load transfer.

In this study, we developed a unified heterogeneous graph representation that simultaneously covers the principal load-bearing components of frame–shear wall structures: shear walls, beams, columns, and shear-wall end columns.

More importantly, the method explicitly incorporates component-level load paths, tracing dead and live loads from slabs through secondary members and shear walls to primary beams, frame columns, and wall-end columns. The model therefore learns not only structural topology, but also how loads are transferred among components and influence their required dimensions.

Using data from 101 real industrial-park projects, the proposed method reduced the average sizing RMSE by approximately 21% compared with the baseline HGNN, with reductions approaching 30% for column dimensions. In three independent engineering cases, more than 70% of the predicted dimensions agreed with engineers’ designs within a ±15% tolerance, over 95% of the components satisfied code requirements without modification, and the overall preliminary design workflow was approximately five times faster.

This work extends our research on intelligent structural design toward a more complete engineering workflow—jointly considering walls, beams, columns, their interactions, and the load paths connecting them.

📄 https://doi.org/10.1016/j.aei.2026.105100

The graphic abstract is generated by Gemini based on the paper.

27/07/2026

🏆 Honored at WAIC 2026: Advancing AI-Driven Building Design with High-Quality Data

We are delighted to share that the Image–Text Topology Dataset for Building Engineering Design was recognized as an Outstanding High-Quality Industry Dataset Achievement at the World Artificial Intelligence Conference (WAIC) 2026 in Shanghai.

Jointly developed by the China Southwest Architectural Design and Research Institute, Tsinghua University, Southwest Jiaotong University, and Hemu AI-Structure, the dataset connects engineering drawings, textual information, and spatial-topological relationships to support AI applications in building design.

The dataset has already contributed to:
- AI-assisted structural design, fire safety design, and building equipment layout
- The development of more than 20 domain-specific AI models
- The continuous improvement of AI-structure Copilot (ai-structure.com), which has supported over 5,200 real-world engineering projects

This recognition at WAIC highlights a fundamental principle: reliable engineering AI depends not only on advanced models, but also on high-quality, domain-grounded data.

We look forward to working with researchers, engineering firms, and technology partners worldwide to accelerate the transition from experience-driven design to data-driven and AI-assisted engineering.

20/07/2026

I just came across this newly published review paper, “Automatic structural design of RC building layout based on deep learning,” and found it very interesting.

Abstract: Driven by rapid urbanization, the demand for high-rise structures, particularly Reinforced Concrete (RC) buildings, continues to rise. In response, numerous methods that automate structural design have recently emerged, aiming for efficient processes at lower cost. Due to the challenges these approaches present, we critically review publications from 2021 to 2025 on the conceptual layout stage of RC buildings, proposing a unified taxonomy that organizes methods by modeled objects (walls, beams, columns), data representations (raster, polygonal, graph), and their targets (e.g., wall/beam length, thickness, position). We analyze biases and constraints in the data and models, discuss how approaches circumvent them, and summarize recurring formulations in tables to help practitioners shortlist candidates. Persistent gaps include limited access to datasets, inconsistent evaluation protocols/metrics, and restricted generalization across plan styles and regulatory contexts. We conclude with guidance on method selection and opportunities to establish standard benchmarks. Although these methods remain largely at the proof-of-concept stage, with current evidence concentrated in high-seismicity shear-wall regimes and cross-code generalization still untested, physics-informed objectives, novel data representations, and cloud computing represent the most promising avenues toward reliable and ultimately scalable automation of RC building design.
Keywords: Conceptual layout; Intelligent design; Structural design; Generative AI; Graph neural networks

Paper: https://doi.org/10.1016/j.jobe.2026.116764

Note: The graphical abstract was generated by Gemini based on the paper.

10/07/2026

Earthquake emergency response at the city scale faces a fundamental data challenge: simulations can generate large datasets, but real building-response measurements are scarce; measured records are valuable, but they are too limited to train large models from scratch.

I am pleased to share our recent paper in Engineering Structures:
A Knowledge Distillation-Based Transfer Learning Framework for Peak Seismic Response Prediction of Urban Building Clusters

In this work, we explored how large-scale simulated seismic response data and limited field measurements can be integrated for rapid, localized prediction of building responses after earthquakes.

The key idea is to use knowledge distillation not merely as model compression, but as a mechanism for simulation-to-measurement knowledge transfer. A large teacher model first learns broad seismic response patterns from a source-domain database containing approximately 9.56 million simulated building-response records. The useful target-relevant knowledge is then distilled into a lightweight student model, which is further fine-tuned using 218 measured building-response records from CESMD.

This “distill-then-transfer” strategy addresses two practical difficulties at the same time. Directly training a small model on measured records suffers from data scarcity. Directly fine-tuning a large simulation-trained model can carry source-domain bias into the target domain. Knowledge distillation provides an intermediate path: it filters, compresses, and re-expresses the simulation-learned knowledge into a compact model that can be more effectively localized using limited observations.

The resulting model achieved R² > 0.89 after transfer, with a 25.3% relative improvement in R² and an approximately 64.5% reduction in mean squared error compared with the non-transferred model. The student model contains only 6,721 parameters, compared with approximately 10⁶ parameters in the teacher model.

More importantly, the framework is designed for continuous localization. As more measured records become available in a target city or region, the model can be updated to better reflect local building inventories, regional construction practices, and recorded earthquake responses. This makes it suitable for rapid post-earthquake screening, regional seismic risk mapping, and intelligent urban emergency response.

The broader message is that AI for earthquake engineering should not rely only on either simulation or observation. Large-scale simulations provide broad physical and structural-response priors, while measured data provide reality-based calibration. By combining the two through knowledge distillation and transfer learning, we can move toward city-scale seismic response models that are both efficient and locally adaptable.

Paper DOI: https://doi.org/10.1016/j.engstruct.2026.123348
Project/GitHub link: https://github.com/Qing.../KD-TL-Seismic-Response-Prediction

08/07/2026

Generative AI for engineering design should not stop at producing geometrically plausible solutions. For real infrastructure, generated designs must also satisfy physical principles, engineering rules, and code-based constraints.

I am pleased to share our recent paper in Advanced Engineering Informatics:
Mechanics-Embedded Diffusion with Closed-Loop Denoising for Intelligent Zoning Design of Concrete-Faced Rockfill Dams

In this work, we explored how diffusion models can be combined with mechanics-based constraints to shift generative design from pattern imitation toward safety-aware, code-compliant engineering design.

The key idea is to embed mechanics into both stages of generation:

During training, a differentiable mechanical loss guides the diffusion model to learn safety-aware design priors. During inference, a surrogate mechanical model evaluates the generated layout step by step and feeds the resulting mechanical deviation back into the denoising process. In this way, the model can dynamically correct the design during generation, rather than simply producing a layout and checking it afterward.

We demonstrate this framework through the zoning design of concrete-faced rockfill dams, in which anti-sliding stability and material zoning are central engineering concerns. The proposed CFRD-Mech-Diffusion framework can generate zoning layouts that satisfy the code-specified safety requirements while improving material allocation and cost efficiency.

In the test case, the method reduced the shares of high-cost zones by 10.17% in the primary rockfill zone and by 4.87% in the modulus-increased zone, while maintaining mechanical compliance and generative accuracy.

The broader message is important: generative AI in engineering should not be only data-driven or visually plausible. It should be science-informed. By integrating physical principles, engineering rules, and design constraints directly into the generative process, AI can become a more reliable partner for engineering design, checking, and optimization.

Although this paper focuses on concrete-faced rockfill dams, the framework is not limited to dams. Similar ideas can be extended to structural and building design problems where geometry, physical performance, code requirements, and decision constraints must be considered together.

This is one step toward generative AI that can not only create designs, but also understand and respect engineering constraints.

Paper: https://doi.org/10.1016/j.aei.2026.105020

02/06/2026

AI-structure Copilot Reaches 5,000+ Shear Wall Design Cases
AI-structure Copilot (https://ai-structure.com/) has now reached 5,000+ shear wall design cases.

For us, this milestone is more than just a number. It means that AI-structure Copilot has been continuously tested and improved through real engineering use, with feedback from structural engineers gradually shaping the product into something more practical, reliable, and aligned with everyday design workflows.

In recent iterations, we focused on making intelligent shear wall design not only capable of generating results, but also easier to review, adjust, and trust in real projects. Key improvements include:
(1) Added settings for bottom strengthening zones
(2) Improved input logic for design conditions
(3) Introduced virtual beams to make slab regions more regular
(4) Optimized batch editing of component dimensions
(5) Improved load modification and component editing after intelligent generation
(6) Enhanced modeling and analysis feedback, so users can better understand problems and respond faster

For structural engineers, efficiency is not only about producing a result. It is also about whether that result fits engineering practice, can be adjusted smoothly, connects with the modeling and analysis workflow, and helps save time in real design work.

That is the direction we will continue to pursue: making AI-structure Copilot more stable, more practical, more transparent, and more useful for structural engineers.

Thanks to everyone who has used the product, shared feedback, and helped us improve it through real engineering cases.

Research Group of Xinzheng Lu

19/08/2026

14/08/2026

12/08/2026

10/08/2026

30/07/2026

27/07/2026

20/07/2026

10/07/2026

08/07/2026

02/06/2026

Address

Website

Alerts

Shortcuts

Share

Category

Listing takedown request