AI bridges the gap in structural engineering
For decades, engineers built bridges the hard way. They drew diagrams, wrote equations, and ran simulations that took hours—sometimes days. Every beam, every load, every vibration had to be calculated by hand or with software that followed rigid, pre-programmed rules. The finite element method, the gold standard for structural analysis, was reliable but slow. It was also expensive. When a bridge designer wanted to test a new shape or material, the computer would churn through millions of calculations, often requiring specialized hardware and days of waiting. The bottleneck was not imagination—it was computation.
Now, something has shifted. Researchers at Tsinghua University and Fuzhou University have published a comprehensive review of how frontier AI techniques are changing this landscape. [2] The review, published in Smart Construction, covers developments from 2020 to 2025 and focuses on three key approaches: graph neural networks, sequence-to-sequence and Transformer-based architectures, and physics-informed methods. The message is clear: the old ways are being supplemented, sometimes replaced, by something faster and more adaptive. But the question remains—what exactly is being lost when we hand over the calculations to machines that think differently than we do?
The Shape of a Problem
The first challenge in computational structural analysis is representation. A building is not a neat grid of numbers. It is a complex arrangement of beams, columns, and joints, each with its own properties and relationships. Traditional finite element methods handle this by breaking the structure into tiny pieces—elements—and solving equations for each one. This works, but it is computationally expensive. The more complex the shape, the more elements you need, and the longer the simulation takes.
Graph neural networks offer a different approach. Instead of breaking the structure into pieces, they treat it as a graph: nodes represent physical points, and edges represent connections between them. This captures the topology of the structure more naturally. Associate Professor Linghan Song from Fuzhou University explains why this matters: “Standard AI often struggles with the complex shapes of buildings. [3] Graph learning, however, could capture the topological information of the structural systems and their irregular meshes more efficiently.” The key word here is “could"—the technology is promising, but it is not yet proven at scale.
The review elaborates on various graph representation methods—how physical structures can be effectively mapped into the digital world as graph data. This is not trivial. A steel beam is not just a line; it has thickness, material properties, and connections that affect its behavior. Mapping these into a graph requires careful design. The researchers discuss model variants and basic concepts, providing a guide for engineers who want to adopt these methods. But they also note the limitations: graph neural networks are still less accurate than finite element methods for some problems, and their behavior is harder to predict.
Learning Sequences, Solving Dynamics
The second frontier is sequence-to-sequence models and Transformers. These architectures, which form the backbone of large language models like GPT, are being repurposed for structural dynamics. Buildings move. They sway in the wind, shake during earthquakes, and settle over time. Simulating these movements requires handling long-range dependencies—how a force applied at one point affects a distant point seconds later.
Chen Wang, a key team member who has successfully developed related algorithms for seismic analysis of building structures, offers insight into how these models work. “Technically speaking, Transformers inherit the architecture of Seq2Seq models,” he explains. “This model family deals with sequential data and is capable of simulating complex engineering behaviors such as long-range dependencies, which are defined as dynamic feature learning.” In plain language: these models can learn how a building responds to forces over time, without being explicitly programmed with the physics equations.
This is both exciting and unsettling. The models learn from data—from simulations or real-world measurements—rather than from first principles. They can be faster than traditional methods once trained, but they require large amounts of high-quality data. And they can make mistakes in ways that are hard to anticipate. An engineer who trusts a Transformer-based model might be surprised when it fails on an unusual input—a building shape it has never seen, or an earthquake pattern that differs from its training data.
The Black Box Problem
This brings us to the central tension in frontier AI for structural engineering: interpretability. Engineers know how to use these models, but they do not always know why they work. Chen Wang puts it bluntly: “Engineers often know how to use AI models, but they aren’t always sure why they work. In structural engineering, this ‘black box’ nature can pose risks to safety and reliability.”
This is not a theoretical concern. Bridges collapse. Buildings fall. When they do, engineers need to understand why—not just to assign blame, but to prevent future failures. If the analysis was done by a black box, the cause may be impossible to determine. The model might have made a subtle error in its internal calculations, or it might have been trained on data that did not include the specific conditions that led to the failure. Either way, the engineer cannot trace the reasoning.
The review shines a spotlight on the rapid evolution of physics-informed methods as a way to bridge this gap. These cutting-edge approaches aim to make AI “understand” the laws of physics. Current strategies include integrating governing equations directly into the neural network’s loss function, forcing the model to respect physical constraints. This does not eliminate the black box entirely, but it constrains its outputs to physically plausible values. If the model predicts a stress that violates the laws of mechanics, it is penalized during training.
Teaching Physics to Neural Networks
The concept of physics-informed neural networks is not new, but it is gaining traction. A separate development from Wuhan University illustrates the potential. Researchers there have developed a transfer learning-enhanced physics-informed neural network for predicting melt pool morphology in selective laser melting—a manufacturing process that uses lasers to melt metal powder layer by layer. This is a different application—selective laser melting in additive manufacturing—but the underlying idea is the same: combine physics constraints with deep learning to achieve better accuracy with less data
Professor Yaowu Hu explains the significance: “This method represents a significant advancement in additive manufacturing. By integrating physics-informed modeling with transfer learning, TLE-PINN bridges the gap between traditional numerical simulations and artificial intelligence, offering precise and efficient solutions for predicting melt pool morphology.” The key innovation is the transfer learning framework, which allows the model to be trained on one set of conditions and then fine-tuned for another. This reduces the amount of data needed for each new application.
The enhanced physics-informed neural network component enforces strong physical constraints during training by incorporating heat transfer equations and boundary conditions directly into the neural network’s loss function. This ensures that the model accurately represents the melt pool morphology, even in complex scenarios. The results are impressive: the model achieves superior accuracy, faster training times, and reduced computational demands compared to traditional methods. But the approach is still experimental, and its generalizability to other domains—like structural engineering—remains to be proven.
The Efficiency Trap
There is a seductive logic to using AI for structural analysis. Faster simulations mean more design iterations. More iterations mean better designs. Better designs mean safer, cheaper buildings. The computational bottleneck that has constrained engineering for decades seems to be dissolving. But there is a trap hidden in this efficiency.
The trap is over-reliance. When simulations become cheap and fast, engineers may run fewer checks, think less about the underlying physics, and trust the model’s outputs without question. This is not a hypothetical risk—it has happened before in other fields. In finance, complex models were used to price derivatives, and when the models failed, the consequences were catastrophic. In medicine, AI diagnostic tools have been shown to perform well on benchmark datasets but fail on real-world patients from different demographics.
The review from Tsinghua and Fuzhou universities acknowledges this risk indirectly. The researchers note that traditional analysis in engineering primarily depends on numerical approaches represented by finite element methods, which can be time-consuming and computationally expensive. Recent advances in frontier AI techniques have shown promising potential to overcome these limitations. But they also emphasize the need for careful validation and physical interpretability. The models are tools, not replacements for engineering judgment.
A Historical Perspective
To understand where we are, it helps to look back at how computational physics has evolved. In the 1960s, finite element methods were developed as a way to approximate solutions to partial differential equations. They were slow and required specialized knowledge to use. Over the following decades, they became faster and more accessible, but the fundamental approach remained the same: discretize the problem, solve the equations, check the results.
In the 1990s, machine learning began to appear in engineering applications, but it was limited to simple tasks like classification and regression. Neural networks were shallow and required careful feature engineering. The idea of using them to solve physics problems directly seemed far-fetched.
The breakthrough came with deep learning in the 2010s. Suddenly, neural networks could learn complex patterns from raw data. Researchers began to explore their potential for scientific computing. The lecture notes from the University of Southern California, compiled for a course developed and taught there, aim to introduce engineering graduate students with a strong background in applied mathematics to select topics in deep learning. The objective is to introduce students familiar with linear algebra and partial differential equations to select topics in deep learning.
The notes exploit the strong connections between deep learning algorithms and conventional techniques of computational physics to achieve two goals: first, to use concepts from computational physics to develop an understanding of deep learning algorithms, and second, to apply novel deep learning algorithms to solve challenging problems in computational physics Not surprisingly, many concepts in deep learning can be connected to similar concepts in computational physics, and one can utilize this connection to better understand these algorithms. Second, several novel deep learning algorithms can be used to solve challenging problems in computational physics. Thus, they offer someone interested in modeling physical phenomena a complementary set of tools.
This is the promise: a two-way street between physics and deep learning. Physics helps us understand deep learning, and deep learning helps us solve physics problems. But the street is still under construction. The connections are there, but they are not always clear, and the traffic is still light.
Where AI Lifts
There are clear areas where AI lifts computational physics. The most obvious is speed. Traditional finite element simulations can take hours or days for complex structures. AI models, once trained, can produce results in seconds or minutes. This is not just a convenience—it enables new types of analysis that were previously impractical. For example, engineers can run thousands of simulations to explore the design space, optimizing for weight, cost, or safety in ways that were not possible before.
Another area where AI lifts is handling complex geometries. Buildings are not simple boxes. They have curves, cantilevers, and irregular shapes that are difficult to mesh with finite elements. Graph neural networks handle these shapes more naturally because they do not require a regular grid. They can work directly with the structure’s topology, capturing relationships that might be missed by traditional methods.
AI also lifts in the area of uncertainty quantification. Engineering is full of unknowns: material properties vary, loads are unpredictable, and construction introduces imperfections. Traditional methods handle this through Monte Carlo simulations, which are computationally expensive. AI models can be trained to approximate the distribution of outcomes, providing probabilistic estimates with much less computation.
Where AI Deceives
But AI also deceives. The most insidious form of deception is the illusion of accuracy. A neural network can produce a precise-looking output—a stress value to five decimal places—that is completely wrong. This happens because neural networks are interpolators, not extrapolators. They are good at predicting outcomes that are similar to their training data, but they can fail catastrophically on inputs that are even slightly different.
This is a particular problem in structural engineering, where failures are rare but catastrophic. A model that has been trained on thousands of simulations of normal buildings may perform perfectly on those cases, but fail on a building with an unusual shape or loading condition. The engineer, lulled by the model’s performance on routine cases, may not realize that the output is unreliable.
Another form of deception is the black box problem. Even when the model is correct, the engineer cannot explain why. This is a problem for certification and liability. Building codes require that designs be justified by calculations that can be reviewed and understood. If the calculation was done by a neural network, how do you review it? How do you certify that it is correct? The review from Tsinghua and Fuzhou universities notes that this “black box” nature can pose risks to safety and reliability.
Where AI Makes Us Superfluous
The third question is where AI makes us superfluous. This is the most uncomfortable question, but it deserves an honest answer. In some areas, AI is not just supplementing human engineers—it is replacing them.
Consider the task of design optimization. Traditionally, an engineer would propose a design, analyze it, and iteratively refine it based on the results. This process requires both domain knowledge and creativity. But AI can now generate designs directly. Generative design algorithms can explore millions of possibilities and produce optimized structures that no human would have thought of. The engineer’s role shifts from creator to curator—selecting among the AI’s proposals rather than generating them from scratch.
This is not necessarily a bad thing. AI-generated designs can be lighter, stronger, and cheaper than human designs. But it raises questions about the future of the engineering profession. If the AI does the creative work, what is left for the engineer? The answer, for now, is judgment. The engineer must evaluate the AI’s proposals, check them for feasibility, and ensure they meet regulatory requirements. But as AI becomes more reliable, even this role may diminish.
The transfer learning-enhanced physics-informed neural network from Wuhan University illustrates this trend. The model predicts melt pool morphology with superior accuracy, faster training times, and reduced computational demands. It does not just assist the engineer—it performs a task that previously required extensive simulation and expertise. The engineer’s role becomes one of oversight, not execution.
The Missing Piece
What is missing from all of this is a clear understanding of when to trust the AI and when to doubt it. The review from Tsinghua and Fuzhou universities provides a valuable guide to the current state of the art, but it does not answer the fundamental question: how do we know when the AI is right?
This is not a technical problem—it is a cultural one. Engineers are trained to verify their calculations, to check assumptions, and to maintain a healthy skepticism. But AI models are opaque. You cannot check their work in the same way you would check a hand calculation or a finite element simulation. The only way to verify an AI model is to test it on a wide range of inputs, but this is expensive and time-consuming, and it can never be exhaustive.
The historical parallel is instructive. When finite element methods were first introduced, many engineers were skeptical. They preferred hand calculations and simple models that they could understand. Over time, the methods were validated, standards were developed, and trust was established. The same process will need to happen for AI. But it will be harder, because the models are more complex and less transparent.
The Same Number, Different Contexts
Throughout this story, a number keeps appearing: 2020 to 2025. This is the period covered by the review from Tsinghua and Fuzhou universities. It is also the period during which the transfer learning-enhanced physics-informed neural network was developed. And it is the period during which the lecture notes from the University of Southern California were being used to train a new generation of engineers.
The number is not arbitrary. This five-year window represents a moment of transition. The techniques that were experimental in 2020 are becoming practical in 2025. The engineers who were students in 2020 are now professionals, equipped with a new set of tools. The question is whether they will use these tools wisely.
Another number appears in the lecture notes: 3 January 2023. This is the date the notes were submitted to arXiv. It is a reminder of how fast this field is moving. The notes were written for a course at the University of Southern California, aimed at engineering graduate students with a strong background in applied mathematics. They are now two years old, which in AI years is ancient. The techniques described in the notes have already been superseded by newer methods.
The Same Phrase, Different Meanings
The phrase “physics-informed” appears in multiple contexts. In the review from Tsinghua and Fuzhou universities, it refers to methods that integrate governing equations into neural networks. In the Wuhan University research, it refers to a specific architecture called the transfer learning-enhanced physics-informed neural network (TLE-PINN). In the lecture notes from the University of Southern California, it is a theme that runs throughout the course.
The phrase means different things in different contexts, but the core idea is the same: use physics to constrain and guide the learning process. This is a response to the black box problem. By forcing the model to respect physical laws, we reduce the risk of it producing physically implausible outputs. But we also limit its flexibility. A physics-informed model cannot learn patterns that violate the laws of physics, even if those patterns are present in the data.
This is both a strength and a weakness. It is a strength because it ensures that the model’s outputs are physically reasonable. It is a weakness because it assumes that the physics is correct and complete. If the governing equations are wrong—if they omit some important effect or make an incorrect assumption—then the physics-informed model will be wrong too.
The Same Question, Different Answers
The question of trust appears in every context. In the review from Tsinghua and Fuzhou universities, it is expressed as a concern about physical interpretability. In the Wuhan University research, it is addressed through the use of physics constraints. In the lecture notes from the University of Southern California, it is implicit in the pedagogical approach—the notes aim to help students understand deep learning by connecting it to familiar concepts from computational physics.
The answers are different, but they all point in the same direction: trust must be earned, not assumed. The review provides a guide for engineers to understand the fundamental concepts, current research status, existing challenges, and future application prospects. The Wuhan University research demonstrates a specific method for achieving trust through physics constraints. The lecture notes provide a foundation for future engineers to build their own understanding.
But none of these answers is complete. Trust in AI for structural engineering will require more than technical solutions. It will require standards, regulations, and a culture of skepticism. It will require engineers to be trained not just in how to use AI, but in how to question it.
The Comparison to Fear
There is a comparison that haunts this story: the comparison to self-driving cars. When autonomous vehicles were first introduced, the promise was that they would be safer than human drivers. The reality has been more complicated. Self-driving cars have been involved in accidents, some fatal, and the reasons for their failures are often opaque. The response has been a mix of regulation, public skepticism, and technological refinement.
Structural engineering is different from driving in important ways. A bridge does not have to react in real time to changing conditions. It can be analyzed, tested, and inspected before it is built. The consequences of failure are catastrophic but rare. And the engineering profession has a long history of learning from failures and improving standards.
But the comparison is still useful. Both self-driving cars and AI-driven structural analysis face the same fundamental challenge: how to trust a system that we do not fully understand. The answer, in both cases, is not to reject the technology, but to approach it with caution, to validate it rigorously, and to maintain human oversight.
The difference is that structural engineering has more time. A bridge can be designed over months or years, not minutes. This gives engineers the opportunity to check the AI’s work, to run independent simulations, and to build confidence in the results. The risk is not that the AI will make a mistake—it is that engineers will stop checking.
The Future That Is Already Here
The review from Tsinghua and Fuzhou universities is not a prediction of the future. It is a description of the present. The techniques it covers—graph neural networks, Transformers, physics-informed methods—are already being used in research labs and some engineering firms. The question is not whether they will become mainstream, but how quickly and with what safeguards.
The computational bottleneck that has constrained engineering for decades is being dissolved. The bottleneck is now human: our ability to understand, validate, and trust the tools we use. The lecture notes from the University of Southern California are a step toward addressing this bottleneck, by training engineers who understand both the physics and the algorithms. But it will take more than education. It will take a cultural shift in the engineering profession.
The Wuhan University research shows what is possible when physics and AI are combined. The transfer learning-enhanced physics-informed neural network achieves results that were previously impossible with either approach alone. But it is still a research prototype, not a production tool. The path from prototype to practice is long and uncertain.
The Last Question
There is one question that the review does not answer, and it is the most important one: when the AI makes a mistake, who is responsible? In traditional engineering, the answer is clear: the engineer who signed off on the design. But if the design was generated by an AI, and the engineer did not understand the AI’s reasoning, who bears the blame?
This is not a hypothetical question. As AI becomes more capable, it will inevitably make mistakes that human engineers would not have made. The question is whether the profession can learn from these mistakes, as it has learned from structural failures in the past.
The review from Tsinghua and Fuzhou universities provides a map of the current landscape, but it does not provide a compass. Engineers must decide how much to trust the AI, when to override it, and how to ensure that the structures they design are safe and reliable.
The bridge that was not there is now being built. The question is whether we can cross it safely, and that answer depends on the engineering profession’s ability to balance innovation with rigorous validation and human oversight.
Sources
1. Science