Some people say that Civil Engineers are one of the oldest professions, and this is not without reason. Since the dawn of humanity, we have needed houses to shelter from wind and rain. Many astonishing buildings have left an indelible mark on human history.
Today, with the rapid development of cities worldwide, more and more skyscrapers are rising. At the same time, with the advancement of engineering, our understanding of building structures and designs is constantly updating. In the past, our ancestors didn’t even know what mechanics was; they explored continuously based on experience. Now, we have advanced theories and technologies to challenge increasingly tall and complex buildings that decorate the urban skyline.
During this development process, many interesting cross-disciplinary fields have emerged, such as the application of Artificial Intelligence (AI) in structural design, which has become a significant trend. AI can help architects and engineers reduce their workload and complete tasks more efficiently.
Recently, Professor Lu Xinzhen from the Department of Civil Engineering at Tsinghua University, recipient of the 2019 “Scientific Exploration Award” in the field of transportation and construction, led a team to develop a Physics-Enhanced Generative Adversarial Network StructGAN-PHY and applied it to structural design.
The research team introduced a physical evaluator into the generative adversarial network, which mainly functions to evaluate the mechanical performance of building structures, thus creating this unique AI that “understands mechanics.” This not only leads to more complete design solutions but can also greatly accelerate the design process, even being 90 times faster than a qualified engineer. The research paper has been published in Earthquake Engineering & Structural Dynamics.
Training AI
The biggest difference between AI and traditional programming is that we don’t have to artificially set many rules for AI; instead, we can let it “play freely” to find the optimal solution. This process often brings many surprises, which is the charm of AI.
All AI relies on a process known as training. In the words of some AI experts, training AI is more like “teaching students” compared to traditional programming. What we need to do is provide it with some datasets, which are a “library of materials” of existing cases, and then use some basic instructions to tell it what the goal is. After that, we just hand the stage over to it.
AI typically starts with “guessing” and then repeatedly studies the dataset, making self-adjustments to discover rules and connections from the data. With proper training, AI has the ability to discover a large number of rules, many of which programmers or anyone else may not even know. In other words, AI can fully self-learn and establish its own knowledge system.
A completed trained AI is like a confident top student. It can then quickly apply what it has learned to specific scenarios, obtaining corresponding answers based on new data it has never seen before.
And all of this is driven by data. It should be noted that a conversational AI does not truly understand human language, and an AI that recognizes cat photos does not genuinely know what a cat is. Instead, AI identifies patterns from vast amounts of trial and error data and applies them.
This naturally brings some problems. A commonly mentioned “counterexample” is when a group of computer scientists intended to train an AI to recognize skin cancer photos, but inadvertently ended up with a ruler AI. This happened because a measuring scale appeared next to the tumors in the provided photo dataset.
It might not be a big deal if an AI “calls a deer a horse,” but for a more professional task like structural design, the consequences could become very serious. Especially if the quality or quantity of the training data is not ideal, the problem may become more severe.
For example, in a structure, there is a type known as shear wall, which primarily bears vertical and lateral forces acting on the building to prevent damage, such as enduring strong winds or the forces brought by seismic shaking. If the design plan has too few shear walls, it naturally brings significant safety hazards; whereas having too many shear walls affects indoor usable space and is not economical.
Therefore, to design a sufficiently safe and economically reasonable building structure, one must not only adhere to the most basic principles of mechanics but also comply with stricter requirements of building structure design codes. Often, AI cannot rely solely on data to cope with all of this.
In this case, we have to let it “learn” some necessary knowledge of mechanics. This is also the direction that Professor Lu Xinzhen and his team are striving for.
Data-Driven Generative Adversarial Network
The newly developed StructGAN-PHY includes several key components. One of the core components is called Generative Adversarial Network (GAN).
GAN is a relatively new type of AI, and the term “adversarial” aptly reflects the characteristics of this type of AI.
GAN actually consists of two neural networks, which are two sets of algorithms running simultaneously, known as the generator (G) and the discriminator (D). During training, they act like a pair of brothers sparring with each other, learning from each other through repeated encounters and improving together.
The generator carefully observes and mimics the input dataset, searching for rules to generate its own “works.” The discriminator is responsible for scoring the generator; it needs to determine the difference between the results obtained by the generator and the real data to make judgments.
In other words, GAN is both an athlete and a referee. Their ultimate common goal is to train the generator into a highly efficient output “master” with outstanding works.
Illustration of the GAN training process.
In this research, the team collected about 150 sets of architectural and structural drawings, along with corresponding design conditions, from more than 10 well-known architectural and structural design institutes in China. Among them, 17 sets of data included detailed structural design models. After preprocessing, this data became the training dataset.
At the beginning of the training, both the generator and the discriminator were like blank slates. The generator might only “scribble” simply, and it may fool the discriminator’s eyes with drawings that only slightly resemble real ones. But through training and continuous trial and error, they ultimately accumulated enough experience. The shear wall design drawings generated by the generator became increasingly “decent,” and the discriminator’s judgment ability also improved.
However, as mentioned earlier, GAN is a data-driven product. For this AI that can design shear walls, it does not understand what a shear wall is, is unclear about building standards, and is completely ignorant of physical principles. What it does is explore rules and patterns from data.
But this is clearly not enough.
Physics-Driven Physical Evaluator
This brings us to the second key part of the research, which is also the key innovation breakthrough of this study, known as the Physical Evaluator.
To enable this AI to make more complete designs, the team hopes to “teach” it some necessary physical knowledge. Or more accurately, they at least need to introduce a method during training to directly inform the generator whether the design schemes it generates are good enough from a physical perspective.
The researchers developed a module called Mechanical Performance Calculator based on previous accumulations. We can understand this calculator as an examiner; when it receives a structural design drawing, it can quickly analyze and calculate the corresponding mechanical parameters based on the structural layout, which are some control indicators. It’s like giving each design drawing a corresponding score.
In the study, the authors chose inter-story displacement angle as a mechanical control indicator as an example.
Illustration of the training process of the physical evaluator.
In this way, we still train by providing a design drawing and its corresponding “score” (label) to the physical evaluator, allowing it to learn to evaluate structural design from a mechanical perspective.
Training is only part of it. Another tricky problem in the research is that scientists must ensure that this evaluation system seamlessly integrates with GAN. In other words, we need to communicate these evaluations in a language that GAN “understands”.
Professor Lu and his team adopted a clever approach by finding a deep neural network to construct a surrogate model. We can understand this surrogate model as being able to “translate” evaluation results relatively accurately to GAN. After comparing the performance of different neural networks, a neural network named ResNet18 (Residual Network) became a qualified “translator”.
After training, ResNet18 replaced the mechanical model and became a physical evaluator that was embedded to work in collaboration with GAN, participating in the training together. The discriminator of GAN is responsible for optimizing the quality of image generation, while the physical evaluator is responsible for optimizing the mechanical performance of the design.
Illustration of the training process of StructGAN-PHY.
After much effort, the team finally developed this physics-enhanced intelligent design method for shear wall structures, which is the complete StructGAN-PHY model.
Physics-Data Coupling Driven
The trained generator has learned well. It is like a real structural engineer, capable of generating corresponding structural design drawings based on necessary input information, such as building drawings and design conditions, while meeting physical design requirements.
Illustration of the application process of StructGAN-PHY.
In the study, the team also conducted case analyses, testing the capabilities of this model under various design conditions.
The results found that this proposed physics-enhanced GAN can generate shear wall structural designs from building drawings and specified design conditions, often yielding solutions closer to engineers’ designs, clearly outperforming designs driven solely by data.
Especially in the absence of corresponding structural design data, this physics-enhanced method can still effectively complete training. Even with a relatively small amount of data, it can significantly improve the mechanical performance of designs.
Typical case analysis results of StructGAN-PHY. From the above image, it can be seen that the physics-enhanced design is closer to the engineers’ plans, while the data-driven design clearly lacks wall configurations. Moreover, according to the measurement of inter-story displacement angles, the data-driven design also yielded results exceeding the code limits.
Overall, StrutGAN-PHY is 44% better than merely data-driven design methods. Its efficiency is also astonishing, with preliminary evaluations showing it is nearly 90 times faster than a qualified engineer.
Combining Computers and Humans
Although this AI, which incorporates mechanics, shows tremendous potential, researchers indicate that it is primarily recommended for the preliminary design stage and still requires further processing, such as drawing more detailed designs based on it before it can truly serve as a complete structural design outcome. Additionally, the errors that may occur during this intelligent design process still need human oversight and verification.
In the future, combining StructGAN-PHY with more AI algorithms may allow AI to better “understand” mechanics and bring more complex mechanical constraints, further optimizing these generated preliminary designs.
#Creative Team:
Written by: WeChat Official Account “Principles”
Reviewed by: Lu Xinzhen, Professor of Civil Engineering at Tsinghua University, recipient of the 2019 “Scientific Exploration Award” in the field of transportation and construction
#References:
https://onlinelibrary.wiley.com/doi/abs/10.1002/eqe.3632
https://mp.weixin.qq.com/s/qWXS10RWb4tCL1Sg3X2Vcg
[U.S.] Janelle Shane, “You Look Like… I Love You: How AI Works and the Strange Things It Brings to the World,” CITIC Publishing Group, April 2021
#Image Sources:
Cover Image: Pxhere
First Image: Pxhere